Compare commits
138
Commits
| Author | SHA1 | Date | |
|---|---|---|---|
|
|
27497fd6be | ||
|
|
466ce3ca82 | ||
|
|
8b5af5ec65 | ||
|
|
34cbc0c0eb | ||
|
|
52ff4bbff5 | ||
|
|
9e7104e362 | ||
|
|
882ae5368f | ||
|
|
12a424ce23 | ||
|
|
c3230761e8 | ||
|
|
f5104008c2 | ||
|
|
013e8c6d0c | ||
|
|
603bba533b | ||
|
|
e202a8d88b | ||
|
|
c56a31c6bc | ||
|
|
b73b18bf9a | ||
|
|
0f351252b2 | ||
|
|
d7d7411fa0 | ||
|
|
45af7f074c | ||
|
|
70ab118612 | ||
|
|
ba0cdea9d0 | ||
|
|
09500ff764 | ||
|
|
b29ce4a0a6 | ||
|
|
311e598b69 | ||
|
|
413dca770c | ||
|
|
9d9239b11e | ||
|
|
a571c41c61 | ||
|
|
ba1fab9dcf | ||
|
|
5dc9c13a71 | ||
|
|
70821bae5a | ||
|
|
486abf82dd | ||
|
|
19afe3af33 | ||
|
|
113042b7f4 | ||
|
|
f8eebd91cb | ||
|
|
00297fc354 | ||
|
|
5baac082f9 | ||
|
|
611b0aa54e | ||
|
|
110f0dbcf5 | ||
|
|
1688cf0a1e | ||
|
|
13b7c8858a | ||
|
|
b8b6f243d1 | ||
|
|
eaa6e805ce | ||
|
|
05dbb21c58 | ||
|
|
f1f4358dee | ||
|
|
3374c41f07 | ||
|
|
95affe87e5 | ||
|
|
5e86398fcd | ||
|
|
844f0cbe42 | ||
|
|
4d5d7847e5 | ||
|
|
06795feed6 | ||
|
|
a4548a7ee1 | ||
|
|
42216865e2 | ||
|
|
e48bdb2401 | ||
|
|
c67eb4b2f3 | ||
|
|
fafa74b0b8 | ||
|
|
049a39aba3 | ||
|
|
cd26430a61 | ||
|
|
aab010ff17 | ||
|
|
9a7bc32051 | ||
|
|
bfdd8ac6be | ||
|
|
018669fb87 | ||
|
|
fb2de78d62 | ||
|
|
7e8d6c1954 | ||
|
|
d21230c6c3 | ||
|
|
4e98d5bfa1 | ||
|
|
8998269377 | ||
|
|
6e9a227ad2 | ||
|
|
4e7506da5a | ||
|
|
06ad93c77c | ||
|
|
0bb7b34259 | ||
|
|
859c17b49d | ||
|
|
46b26c9624 | ||
|
|
fd215c18e4 | ||
|
|
efa3afc7dc | ||
|
|
e03dd66051 | ||
|
|
3476fe5fae | ||
|
|
3e08c8347e | ||
|
|
c691123d2d | ||
|
|
a4b81535e8 | ||
|
|
a898cf4b65 | ||
|
|
10009a0a43 | ||
|
|
4dfe47ff99 | ||
|
|
bd997e61d6 | ||
|
|
23d0670126 | ||
|
|
12c600edbe | ||
|
|
e4115e705e | ||
|
|
0cbc4a2d72 | ||
|
|
cea6d35729 | ||
|
|
4c56918d8a | ||
|
|
f17263417e | ||
|
|
6f0b463489 | ||
|
|
28c01caa1a | ||
|
|
95b59e9b0a | ||
|
|
bc967f1f6e | ||
|
|
4228b8a74f | ||
|
|
300086fc83 | ||
|
|
688eb262a9 | ||
|
|
bfb6f5dede | ||
|
|
88c0fbe56b | ||
|
|
3d74eff224 | ||
|
|
22d08afe2d | ||
|
|
a97f757e34 | ||
|
|
bbc8f13944 | ||
|
|
f18a40d76c | ||
|
|
c7644a36aa | ||
|
|
f394f31c12 | ||
|
|
0513673c2a | ||
|
|
a73542a391 | ||
|
|
c465797654 | ||
|
|
7ce3f982f2 | ||
|
|
6311456917 | ||
|
|
3c06559c5d | ||
|
|
5c9f2e8c83 | ||
|
|
62902faaca | ||
|
|
d22fd77b22 | ||
|
|
2f86f493d6 | ||
|
|
eafa0fad85 | ||
|
|
2e7c497b69 | ||
|
|
79dcfdcf9a | ||
|
|
487e28a904 | ||
|
|
c492f97a67 | ||
|
|
ec2ff74514 | ||
|
|
c501f8d6e6 | ||
|
|
b4287cbfda | ||
|
|
26724a1db1 | ||
|
|
d6368ee709 | ||
|
|
ea211a5c0b | ||
|
|
9e8b433c95 | ||
|
|
2a1c62aeed | ||
|
|
948d6f4b47 | ||
|
|
725271a7da | ||
|
|
4ea62e6886 | ||
|
|
c01cf3b078 | ||
|
|
c67777a6b9 | ||
|
|
e5e5e7560c | ||
|
|
6e1cd5fbbc | ||
|
|
8bd8f7c0bb | ||
|
|
1c42cd1f78 | ||
|
|
be090a43f2 |
No files matched your search
Vendored
+3
@@ -0,0 +1,3 @@
|
||||
{
|
||||
"CodeGPT.apiKey": "Google AI Studio"
|
||||
}
|
||||
@@ -0,0 +1,465 @@
|
||||
# mindnet v2.2 – Dokumentationsrahmenplan
|
||||
**Datei:** `docs/mindnet_docs_master_plan_v2.2.md`
|
||||
**Stand:** 2025-12-06
|
||||
**Bezug:** Programmplan mindnet v2.2 :contentReference[oaicite:0]{index=0}
|
||||
|
||||
---
|
||||
|
||||
## 1. Zweck dieses Rahmenplans
|
||||
|
||||
Dieser Rahmenplan definiert:
|
||||
|
||||
- die **Zieldokumente** der mindnet v2.2 Dokumentation,
|
||||
- deren **Dateinamen, Zielgruppen, Inhalte und Quellen**,
|
||||
- die **Zuordnung** zu Workpackages (WP) und Release-Tags,
|
||||
- die **Regeln zum Informations-Erhalt** (keine inhaltlichen Verluste),
|
||||
- die **empfohlene Bearbeitungsreihenfolge**.
|
||||
|
||||
Alle hier beschriebenen Zieldokumente werden als **eigene Markdown-Dateien** gepflegt und sind Teil von WP-14 _Review / Refactoring / Dokumentation_ (Phase E). :contentReference[oaicite:1]{index=1}
|
||||
|
||||
---
|
||||
|
||||
## 2. Prinzipien zur Sicherung der Informationen
|
||||
|
||||
Um zu verhindern, dass Inhalte oder Ergebnisse verloren gehen, gelten folgende Prinzipien:
|
||||
|
||||
1. **Programmplan v2.2 als Masterreferenz**
|
||||
- Programmauftrag, Vision, Phasenmodell und WP-Status werden **nicht verändert**, sondern nur referenziert.
|
||||
- `Programmplan_V2.2.md` bleibt das führende Dokument für Ziele, WPs, Ampelstatus. :contentReference[oaicite:2]{index=2}
|
||||
|
||||
2. **Keine unkontrollierte Löschung von Inhalten**
|
||||
- Bestehende Dokumente (`knowledge_design.md`, `chunking_strategy.md`, `mindnet_technical_architecture.md`, `Handbuch.md`, `TYPE_REGISTRY_MANUAL.md`, `docs_mindnet_retriever.md`, `wp04_retriever_scoring.md`) werden **nicht gelöscht**, sondern:
|
||||
- vollständig archiviert,
|
||||
- in die neuen Zieldokumente **überführt**,
|
||||
- oder **explizit** in den Anhängen (I) als „legacy / historisch“ verlinkt.
|
||||
|
||||
3. **Explizite Quellenzuordnung („Building Blocks“)**
|
||||
- Für jede Sektion eines Zieldokuments wird angegeben, aus **welchen bestehenden Dateien und Abschnitten** der Inhalt stammt.
|
||||
- Neue Inhalte (z. B. User Guide H) werden klar als „neu, v2.2“ gekennzeichnet.
|
||||
|
||||
4. **Ist-Stand vs. Roadmap trennen**
|
||||
- Implementierter Stand v2.2: WP01–WP04a (🟢) mit Importer, Chunking, Edge-Pipeline, Typ-Registry und hybrider Retriever.
|
||||
- Geplante Features (Self-Healing, Virtual Schema Layer, Explanation Layer, Feedback-Logging, MCP-Agenten) werden als **Roadmap-Funktionen** kenntlich gemacht (Phase B–D, WPs 05–13).
|
||||
|
||||
5. **Versionierbarkeit**
|
||||
- Alle Zieldokumente tragen im Kopf eine **Version** und **Stand-Datum** sowie optional einen Verweis auf den letzten Gitea-Tag (z. B. `mindnet_v2.2_docs_initial`, `WP04a-Retriever_final`). :contentReference[oaicite:6]{index=6}
|
||||
- Änderungen werden im Abschnitt „Changelog“ der jeweiligen Datei notiert.
|
||||
|
||||
6. **Keine funktionalen Annahmen ohne Verifikation**
|
||||
- Wo das Verhalten nicht eindeutig aus Code-/Handbuch-/WP-Dokumenten hervorgeht, wird der Punkt zunächst als _offene Frage_ markiert und nicht als Tatsache dokumentiert.
|
||||
- Verifizierte Funktionen (z. B. `import_markdown`, Edge-Schema, `types.yaml`, `/query`, Scoring-Formel) werden exakt gemäß der bestehenden Dokumente beschrieben.
|
||||
|
||||
---
|
||||
|
||||
## 3. Zieldokumente der mindnet v2.2 Dokumentation
|
||||
|
||||
### 3.1 Übersichtstabelle
|
||||
|
||||
| ID | Zieldatei | Rolle/Zielgruppe | Hauptquellen (Building Blocks) |
|
||||
|----|------------------------------------------|---------------------------------------------|--------------------------------|
|
||||
| A | `mindnet_overview_v2.2.md` | Einstieg für alle Rollen | Programmplan v2.2, Knowledge Design |
|
||||
| B | `mindnet_functional_architecture_v2.2.md`| Architekt:innen, Mindmaster | Knowledge Design, Functional Architecture, Programmplan |
|
||||
| C | `mindnet_technical_architecture_v2.2.md` | Entwickler:innen, Tech-Architekt:innen | mindnet_technical_architecture, chunking_strategy, Handbuch |
|
||||
| D | `mindnet_knowledge_design_manual_v2.2.md`| Vault-Autor:innen, Mindmaster, Analyst:innen| knowledge_design, TYPE_REGISTRY_MANUAL, Functional Architecture |
|
||||
| E | `mindnet_pipeline_playbook_v2.2.md` | Dev/Ops, Tech-Lead | mindnet_v2_implementation_playbook, Handbuch, chunking_strategy, docs_mindnet_retriever, wp04_retriever_scoring |
|
||||
| F | `mindnet_developer_guide_v2.2.md` | Entwickler:innen (Code, API, Tests) | Technical Architecture, Handbuch, docs_mindnet_retriever |
|
||||
| G | `mindnet_admin_guide_v2.2.md` | Admins, Betreiber | Handbuch, Implementation Playbook, Programmplan (Governance) |
|
||||
| H | `mindnet_user_guide_v2.2.md` | Endnutzer:innen, Agent-User | Knowledge Design, Programmplan (Use Cases), spätere MCP-WPs |
|
||||
| I | `mindnet_appendices_v2.2.md` | Alle Rollen (Referenzen, Tabellen) | TYPE_REGISTRY_MANUAL, chunking_strategy, WP-Übersicht, Edge-Referenz |
|
||||
|
||||
---
|
||||
|
||||
## 4. Detailpläne pro Zieldokument (Inhaltsverzeichnis & Building Blocks)
|
||||
|
||||
### A) `mindnet_overview_v2.2.md`
|
||||
|
||||
#### Zweck & Zielgruppe
|
||||
- Einstieg in mindnet für alle Rollen (Owner, Mindmaster, Developer, Admin, Agenten-Entwickler).
|
||||
- Auf 3–5 Seiten: _Was ist mindnet? Wie ist es gedacht? Wie lese ich die Dokumentation?_ :contentReference[oaicite:17]{index=17}
|
||||
|
||||
#### Geplantes Inhaltsverzeichnis
|
||||
|
||||
1. **Einführung**
|
||||
1. Programmauftrag & Vision (v2.2)
|
||||
2. Was mindnet ist – und was nicht
|
||||
2. **Drei Ebenen von mindnet**
|
||||
1. Content-Ebene (Vault, Notes, Chunks)
|
||||
2. Semantische Ebene (Edges, Typen, Retriever)
|
||||
3. Identitäts- / Persönlichkeits-Ebene (Werte, Entscheidungen, Profile – Roadmap)
|
||||
3. **End-to-End-Überblick**
|
||||
1. Datenfluss: Vault → Import → Qdrant → Retriever → Antwort
|
||||
2. WP-Phasenmodell (A–E) und aktueller Stand (WP01–04a grün)
|
||||
4. **Rollen & Perspektiven**
|
||||
1. Mindmaster / Owner
|
||||
2. Autor:in (Vault-Notes)
|
||||
3. Developer / Architect
|
||||
4. Administrator:in
|
||||
5. Agenten & LLMs (MCP-Perspektive)
|
||||
5. **Leserführung durch die Dokumentation**
|
||||
1. Welche Datei für welche Rolle?
|
||||
2. Wie mit Ist-Stand vs. Roadmap umgehen?
|
||||
6. **Glossar zentraler Begriffe**
|
||||
- mindnet, Vault, Note, Chunk, Edge, Type, Identity, Self-Healing, Self-Tuning, Explanation Layer, Virtual Schema Layer, etc.
|
||||
|
||||
#### Building Blocks (Quellen-Mapping)
|
||||
|
||||
- Programmauftrag & Vision → `Programmplan_V2.2.md`, Kapitel 1–2. :contentReference[oaicite:18]{index=18}
|
||||
- Phasenmodell & WPs → `Programmplan_V2.1.md`/v2.2, Kapitel 5–9 (angepasst auf v2.2-Stand).
|
||||
- High-Level-Architektur → `knowledge_design.md`, Kapitel 1–3 (Ziel, Zielarchitektur, Collections). :contentReference[oaicite:20]{index=20}
|
||||
- Rollenmodell → Ableitung aus Knowledge Design + Programmplan (Owner, Agenten, MCP-WPs).
|
||||
|
||||
---
|
||||
|
||||
### B) `mindnet_functional_architecture_v2.2.md`
|
||||
|
||||
#### Zweck & Zielgruppe
|
||||
- Fachliche Architektur von mindnet (Konzeptmodell).
|
||||
- Für Architekt:innen, Mindmaster, Knowledge-Designer.
|
||||
|
||||
#### Geplantes Inhaltsverzeichnis
|
||||
|
||||
1. **Ziel & Einordnung**
|
||||
1. Bezug zu Programmauftrag & WPs (WP01–WP07, WP10–12)
|
||||
2. **Kernkonzepte**
|
||||
1. Notes & Note-Typen (concept, experience, project, decision, value, principle, person, event, …)
|
||||
2. Chunks & Sections
|
||||
3. Edges (Struktur, explizit, inline, Default-Kanten)
|
||||
4. Typen-Registry & Virtual Schema Layer (fachliche Sicht, Roadmap)
|
||||
3. **Graph-Modell**
|
||||
1. Knotenarten (Notes, Chunks, Identity Nodes, People)
|
||||
2. Kantenarten (belongs_to, references, similar_to, depends_on, related_to, next/prev, …)
|
||||
3. Scope & Provenienz (chunk vs. note, rule_id, confidence)
|
||||
4. **End-to-End fachlicher Datenfluss**
|
||||
1. Erfassung von Wissen (Note erstellen)
|
||||
2. Import & Graphaufbau
|
||||
3. Retrieval & Begründungspfade
|
||||
4. Self-Healing & Retroactive Edges (WP06/07 – Roadmap)
|
||||
5. **Szenarien & Use Cases**
|
||||
1. Wissensspeicherung (z. B. Leitbild, Projekt, Erfahrung)
|
||||
2. Entscheidungen & Why-Layer (WP04b – Roadmap)
|
||||
3. Persönlichkeits-/Identitätsmodell (WP10–12 – Roadmap)
|
||||
6. **Rollenmodell (fachliche Sicht)**
|
||||
1. Mindmaster
|
||||
2. Content-Autor:innen
|
||||
3. Analyst:innen (Graph-Leser)
|
||||
7. **Designprinzipien (fachlich)**
|
||||
- Late Binding, Virtual Schema, Self-Healing, Minimalinvasives Schreiben etc.
|
||||
|
||||
#### Building Blocks
|
||||
|
||||
- `knowledge_design.md` – Kapitel 1–4, 6–7 (Ziel, Architektur, Collections, Identitäten, Prinzipien). :contentReference[oaicite:28]{index=28}
|
||||
- `mindnet_functional_architecture.md` – alle Kapitel (Kernkonzepte, Edgetypen, Provenienz, Rollen). :contentReference[oaicite:29]{index=29}
|
||||
- `chunking_strategy.md` – Edges, Datenmodell, Graph-Sicht. :contentReference[oaicite:30]{index=30}
|
||||
- Programmplan – WPs 01–07, 10–12 für Roadmap-Features.
|
||||
|
||||
---
|
||||
|
||||
### C) `mindnet_technical_architecture_v2.2.md`
|
||||
|
||||
#### Zweck & Zielgruppe
|
||||
- Technische Architektur des Systems: Collections, Payload-Felder, Services, Schemata.
|
||||
- Für Entwickler:innen und Tech-Architekt:innen.
|
||||
|
||||
#### Geplantes Inhaltsverzeichnis
|
||||
|
||||
1. **Ziel & Scope**
|
||||
1. Abgrenzung zur funktionalen Architektur
|
||||
2. Bezug zu WPs 02–04a, 05 (Schema, Importer, Retriever)
|
||||
2. **Systemübersicht**
|
||||
1. Komponentenübersicht (Importer, Qdrant, Retriever-API, Embedding-Service, n8n/MCP-Ausblick)
|
||||
2. Ports, Dienste, Hauptverzeichnisse
|
||||
3. **Datenmodell in Qdrant**
|
||||
1. Notes (`mindnet_notes`)
|
||||
2. Chunks (`mindnet_chunks`)
|
||||
3. Edges (`mindnet_edges`)
|
||||
4. Optionale Collections (`mindnet_people`, zukünftige)
|
||||
4. **Typ-Registry & Konfiguration**
|
||||
1. `config/types.yaml` – Aufbau, Defaults, Typauflösung (Ist-Stand: Frontmatter-Overrides deaktiviert)
|
||||
2. `config/retriever.yaml` – Scoringgewichte, Edge-Typ-Gewichte :contentReference[oaicite:35]{index=35}
|
||||
5. **Import-Pipeline (technische Sicht)**
|
||||
1. Discovery & Parsing (`import_markdown.py`)
|
||||
2. Hash-/Idempotenz-Mechanismus (ENV-Variablen, Baseline-Modi)
|
||||
3. Chunking (`chunker.py`, `chunk_payload.py`)
|
||||
4. Edge-Erzeugung (`derive_edges.py`)
|
||||
5. Roundtrip & Tests (Export & Vault-Vergleich)
|
||||
6. **Retriever-Architektur**
|
||||
1. API (FastAPI-Endpunkt `/query`)
|
||||
2. Semantik- und Hybridmodus
|
||||
3. Subgraph & Graph-API (expand)
|
||||
4. Scoring-Formel und Komponenten (semantik, edges, centrality) :contentReference[oaicite:37]{index=37}
|
||||
7. **Offene technische Punkte / Roadmap**
|
||||
1. Explanation Layer (WP04b)
|
||||
2. Feedback-Logging (WP04c)
|
||||
3. Virtual Schema Layer (WP05)
|
||||
4. Self-Healing & Self-Tuning (WP06–08)
|
||||
|
||||
#### Building Blocks
|
||||
|
||||
- `mindnet_technical_architecture.md` – Kerntext. :contentReference[oaicite:39]{index=39}
|
||||
- `chunking_strategy.md` – Datenmodell & Chunk-Details. :contentReference[oaicite:40]{index=40}
|
||||
- `Handbuch.md` – ENV, Skripte, Runbooks. :contentReference[oaicite:41]{index=41}
|
||||
- `wp04_retriever_scoring.md` / `docs_mindnet_retriever.md` – Retriever-Details. :contentReference[oaicite:42]{index=42}
|
||||
|
||||
---
|
||||
|
||||
### D) `mindnet_knowledge_design_manual_v2.2.md`
|
||||
|
||||
#### Zweck & Zielgruppe
|
||||
- Konkrete Anleitung für das Schreiben und Strukturieren von Notes & Chunks im Vault.
|
||||
- Für dich als Owner/Mindmaster und alle, die Notes erstellen.
|
||||
|
||||
#### Geplantes Inhaltsverzeichnis
|
||||
|
||||
1. **Zweck & Scope**
|
||||
1. Rolle des Vaults als „Source of Truth“
|
||||
2. **Note-Struktur & Frontmatter**
|
||||
1. Pflicht- und optionale Felder (id, title, type, tags, status, created/updated, …)
|
||||
2. Empfehlungen für IDs, Pfade und Namenskonventionen
|
||||
3. **Note-Typen & Typ-Registry**
|
||||
1. Übersicht der Typen (concept, project, experience, decision, value, principle, person, event, …)
|
||||
2. Typ-spezifische Einsatzempfehlungen
|
||||
3. Zusammenspiel mit `types.yaml` (retriever_weight, chunk_profile, edge_defaults)
|
||||
4. **Edges & Referenzen in Notes**
|
||||
1. Wikilinks (`[[Note]]`) → `references`
|
||||
2. Inline-Relationen (`[[rel:depends_on XYZ]]`, `similar_to`, `related_to`)
|
||||
3. Callout-Edges (`> [!edge] …`)
|
||||
4. Strukturkanten (Ordner, belongs_to)
|
||||
5. **Best Practices**
|
||||
1. Beispiele für gut modellierte Notes (Projekt, Entscheidung, Erfahrung)
|
||||
2. Typische Anti-Patterns und wie man sie vermeidet
|
||||
6. **Langfristige Stabilität**
|
||||
1. Virtual Schema Layer – was bedeutet das für Note-Autoren? (Wenig Änderungen in bestehenden Notes, Logik in Configs) :contentReference[oaicite:46]{index=46}
|
||||
|
||||
#### Building Blocks
|
||||
|
||||
- `knowledge_design.md` – Kerntext + Beispiele. :contentReference[oaicite:47]{index=47}
|
||||
- `TYPE_REGISTRY_MANUAL.md` – Typ-Registry & Verhalten. :contentReference[oaicite:48]{index=48}
|
||||
- `mindnet_functional_architecture.md` – ergänzende Konzepte & Edges. :contentReference[oaicite:49]{index=49}
|
||||
- `chunking_strategy.md` – Inline-Relationen, Edges, Datenmodell. :contentReference[oaicite:50]{index=50}
|
||||
|
||||
---
|
||||
|
||||
### E) `mindnet_pipeline_playbook_v2.2.md`
|
||||
|
||||
#### Zweck & Zielgruppe
|
||||
- End-to-End-Playbook für die gesamte Pipeline:
|
||||
- Importer,
|
||||
- Chunking,
|
||||
- Edge-Erzeugung,
|
||||
- Retriever & Scoring,
|
||||
- (perspektivisch) Explanation Layer, Self-Healing, Self-Tuning.
|
||||
- Für Dev/Ops und technische Leads.
|
||||
|
||||
#### Geplantes Inhaltsverzeichnis
|
||||
|
||||
1. **Zweck & Einordnung**
|
||||
1. Bezug zu WPs 02–04a (Ist) und 04b–08 (Roadmap)
|
||||
2. **Import-Pipeline (Runbook-Sicht)**
|
||||
1. Inputs (Vault, Config-Files, ENV)
|
||||
2. Standardläufe (Initialimport, inkrementelle Updates, Sync-Deletes)
|
||||
3. Baseline-Builds & Hash-Modi
|
||||
3. **Chunking & Payload-Aufbau**
|
||||
1. Chunk-Profile, Overlaps, Fenster
|
||||
2. Feldbesetzung (`text`, `window`, `section_title`, `neighbors`, …)
|
||||
4. **Edge-Erzeugung**
|
||||
1. Strukturkanten (belongs_to, next, prev)
|
||||
2. Explizite Edges (Wikilinks, Inline-Relationen, Callouts)
|
||||
3. Typ-Default-Kanten aus `types.yaml`
|
||||
5. **Retriever & Scoring**
|
||||
1. Konfiguration (`retriever.yaml`, ENV)
|
||||
2. Scoring-Formel und Parameter
|
||||
3. Semantik- vs. Hybridmodus
|
||||
4. Test- und Smoke-Suites
|
||||
6. **Quality Gates & Tests**
|
||||
1. Roundtrip-Tests (Vault → Qdrant → Export)
|
||||
2. Chunk-Integrität und Fenster-Checks
|
||||
3. Retriever-Regressionstests
|
||||
7. **Ausblick: Self-Healing, Retroactive Edges, Self-Tuning**
|
||||
1. Geplante Batch-Jobs und Feedback-Pipelines (WP06–08)
|
||||
|
||||
#### Building Blocks
|
||||
|
||||
- `mindnet_v2_implementation_playbook.md` – Policies, Teststrategie, KPIs. :contentReference[oaicite:57]{index=57}
|
||||
- `Handbuch.md` – Skripte & Test-Runbooks. :contentReference[oaicite:58]{index=58}
|
||||
- `chunking_strategy.md` – Chunk- & Edge-Strategie. :contentReference[oaicite:59]{index=59}
|
||||
- `docs_mindnet_retriever.md` / `wp04_retriever_scoring.md` – Retriever-Praktiken & Tests. :contentReference[oaicite:60]{index=60}
|
||||
|
||||
---
|
||||
|
||||
### F) `mindnet_developer_guide_v2.2.md`
|
||||
|
||||
#### Zweck & Zielgruppe
|
||||
- Konzentrierte Sicht für Entwickler:innen:
|
||||
- wie das Projekt strukturiert ist,
|
||||
- wie man es lokal aufsetzt,
|
||||
- wie man neue Features implementiert/testet.
|
||||
|
||||
#### Geplantes Inhaltsverzeichnis
|
||||
|
||||
1. **Projektstruktur**
|
||||
1. Verzeichnisbaum (app/, scripts/, tests/, config/, …)
|
||||
2. **Lokale Entwicklungsumgebung**
|
||||
1. Abhängigkeiten, venv, ENV
|
||||
2. Starten von Services (Qdrant, API, ggf. ollama)
|
||||
3. **APIs & Module**
|
||||
1. Importer-Skripte
|
||||
2. Core-Module (Chunker, Edge-Engine, Retriever)
|
||||
3. FastAPI-Endpunkte (`/query`, evtl. Admin-Endpoints)
|
||||
4. **Tests & Debugging**
|
||||
1. pytest-Suites
|
||||
2. Logging & Diagnoseskripte
|
||||
5. **Guidelines für Erweiterungen**
|
||||
1. Neue Typen & Edge-Regeln
|
||||
2. Änderungen am Datenmodell
|
||||
3. Umgang mit Schema-Änderungen (Virtual Schema Layer)
|
||||
|
||||
#### Building Blocks
|
||||
|
||||
- `mindnet_technical_architecture.md` – Modul- & API-Beschreibungen. :contentReference[oaicite:63]{index=63}
|
||||
- `Handbuch.md` – Skripte & Beispiele. :contentReference[oaicite:64]{index=64}
|
||||
- `docs_mindnet_retriever.md` – API-Beispiele für den Retriever. :contentReference[oaicite:65]{index=65}
|
||||
|
||||
---
|
||||
|
||||
### G) `mindnet_admin_guide_v2.2.md`
|
||||
|
||||
#### Zweck & Zielgruppe
|
||||
- Betriebshandbuch für mindnet:
|
||||
- Installation,
|
||||
- Konfiguration,
|
||||
- Backup/Restore,
|
||||
- Monitoring,
|
||||
- Troubleshooting.
|
||||
|
||||
#### Geplantes Inhaltsverzeichnis
|
||||
|
||||
1. **Zielgruppe & Scope**
|
||||
2. **Initial Setup**
|
||||
1. Voraussetzungen (Ubuntu, Docker/Qdrant, Python-Umgebung)
|
||||
2. ENV-Variablen & Config-Files (types.yaml, retriever.yaml)
|
||||
3. **Betrieb im Alltag**
|
||||
1. Regelmäßige Imports
|
||||
2. Umgang mit Fehlermeldungen im Importer
|
||||
3. Health-Checks für Qdrant & API
|
||||
4. **Update-Prozess**
|
||||
1. Vorgehen bei Schema-/Code-Updates
|
||||
2. Umgang mit neuen Typen und Policies
|
||||
5. **Backup & Restore**
|
||||
1. Qdrant-Snapshots
|
||||
2. Vault-Backup (Obsidian)
|
||||
3. Konsistenzprüfungen
|
||||
6. **Monitoring & Logging**
|
||||
1. typische Logquellen (Importer, API, n8n)
|
||||
7. **Governance & Sicherheit**
|
||||
1. Bezug zur Programm-Governance (WP14, MCP-Sicherheit)
|
||||
|
||||
#### Building Blocks
|
||||
|
||||
- `Handbuch.md` – Betriebsskripte & ENV. :contentReference[oaicite:69]{index=69}
|
||||
- `mindnet_v2_implementation_playbook.md` – Policies/Resilienz. :contentReference[oaicite:70]{index=70}
|
||||
- Programmplan – Governance / WP14. :contentReference[oaicite:71]{index=71}
|
||||
|
||||
---
|
||||
|
||||
### H) `mindnet_user_guide_v2.2.md`
|
||||
|
||||
#### Zweck & Zielgruppe
|
||||
- Anleitung für Nutzer:innen (inkl. späterer Chat-Agenten-User):
|
||||
- wie man mindnet im Alltag nutzt,
|
||||
- wie man Fragen stellt,
|
||||
- wie man neue Gedanken einbringt.
|
||||
|
||||
#### Geplantes Inhaltsverzeichnis
|
||||
|
||||
1. **Was mindnet für dich tut**
|
||||
1. Persönliches Gedächtnis
|
||||
2. KI-Zwilling als Gesprächspartner :contentReference[oaicite:72]{index=72}
|
||||
2. **Zugänge zu mindnet**
|
||||
1. Obsidian-Vault
|
||||
2. Chat-Agent (lokal, n8n-Workflows)
|
||||
3. (später) Interview-Assistent, MCP-Tools
|
||||
3. **Gute Fragen an mindnet**
|
||||
1. Beispiel-Queries (Wissensabruf, Entscheidungen, Reflexionen)
|
||||
4. **Neue Inhalte eingeben**
|
||||
1. Notes in Obsidian erstellen
|
||||
2. Typen und Links bewusst setzen (vereinfachte Version aus D)
|
||||
5. **Ergebnisse interpretieren**
|
||||
1. Was bedeutet ein Treffer-Set?
|
||||
2. Wie sind Scores zu lesen? (nur Ist-Stand, Explanation Layer als Ausblick)
|
||||
6. **Feedback & Lernen**
|
||||
1. Manuelles Feedback (gut/schlecht) – Roadmap WP04c/08
|
||||
|
||||
#### Building Blocks
|
||||
|
||||
- Programmplan – Vision & Use-Cases. :contentReference[oaicite:76]{index=76}
|
||||
- Knowledge Design – Autorensicht (stark vereinfacht). :contentReference[oaicite:77]{index=77}
|
||||
- Retriever-Dokumente – Scoring & Interpretation. :contentReference[oaicite:78]{index=78}
|
||||
|
||||
---
|
||||
|
||||
### I) `mindnet_appendices_v2.2.md`
|
||||
|
||||
#### Zweck & Zielgruppe
|
||||
- Referenzen, Tabellen, technische Details, die Hauptdokumente nicht überfrachten sollen.
|
||||
|
||||
#### Geplantes Inhaltsverzeichnis
|
||||
|
||||
1. **Typenreferenz (Auszug aus `types.yaml`)**
|
||||
- Tabellenform: Typ, chunk_profile, edge_defaults, retriever_weight
|
||||
2. **Edge-Typen & Bedeutungen**
|
||||
- belongs_to, references, depends_on, similar_to, related_to, next, prev, …
|
||||
3. **Datenmodelle (JSON-Skizzen)**
|
||||
- Note, Chunk, Edge-Payload (aktuell implementierter Stand)
|
||||
4. **Relevante ENV-Variablen & Konfigurationen**
|
||||
- Hash-Settings, COLLECTION_PREFIX, MINDNET_TYPES_FILE, MINDNET_RETRIEVER_CONFIG etc.
|
||||
5. **WP-Übersicht & Status**
|
||||
- kompakte Tabelle aller WPs mit Status (🟢/🟡) und Verweisen auf relevante Dokumente.
|
||||
6. **Historische Dokumente**
|
||||
- Verweis auf alte Versionen (`mindnet_functional_architecture.md`, ältere Playbooks), falls diese nicht vollständig eingearbeitet werden.
|
||||
|
||||
#### Building Blocks
|
||||
|
||||
- `TYPE_REGISTRY_MANUAL.md` – Typen-Referenz. :contentReference[oaicite:84]{index=84}
|
||||
- `chunking_strategy.md` – JSON-Beispiele. :contentReference[oaicite:85]{index=85}
|
||||
- `mindnet_technical_architecture.md` – Felder & Prioritäten. :contentReference[oaicite:86]{index=86}
|
||||
- Programmplan – WP-Tabellen.
|
||||
|
||||
---
|
||||
|
||||
## 5. Bearbeitungsreihenfolge (Empfehlung)
|
||||
|
||||
Aus Sicht der Informationssicherung und Konsistenz ist folgende Reihenfolge sinnvoll:
|
||||
|
||||
1. **C – `mindnet_technical_architecture_v2.2.md`**
|
||||
- Begründung:
|
||||
- bildet den **aktuellen implementierten Stand** von WP02–WP04a am präzisesten ab,
|
||||
- ist stark an konkrete Felder, Collections und Skripte gebunden (hier ist der Ist-Stand am eindeutigsten belegbar),
|
||||
- ist die Grundlage für Developer Guide, Playbook und Admin Guide.
|
||||
|
||||
2. **E – `mindnet_pipeline_playbook_v2.2.md`**
|
||||
- Konsolidiert Handbuch + Implementation-Playbook + WP04-Dokumente.
|
||||
|
||||
3. **D – `mindnet_knowledge_design_manual_v2.2.md`**
|
||||
- Stellt sicher, dass das Wissensdesign stabil dokumentiert ist, bevor Overview/User Guide entstehen.
|
||||
|
||||
4. **A – `mindnet_overview_v2.2.md`**
|
||||
- Kann sich dann sauber auf die finalisierte fachliche/technische Sicht stützen.
|
||||
|
||||
5. **F, G, H, I**
|
||||
- Developer/Admin/User Guides & Appendices, basierend auf den konsolidierten Kern-Dokumenten.
|
||||
|
||||
---
|
||||
|
||||
## 6. Nächster konkreter Schritt
|
||||
|
||||
Als nächstes finalisieren wir gemäß dieser Reihenfolge das Dokument:
|
||||
|
||||
> **C – `mindnet_technical_architecture_v2.2.md`**
|
||||
|
||||
Dabei gehen wir wie folgt vor:
|
||||
|
||||
1. Bestehendes `mindnet_technical_architecture.md` **vollständig** als Basis übernehmen. :contentReference[oaicite:89]{index=89}
|
||||
2. Fehlende Details aus `chunking_strategy.md`, `Handbuch.md`, `TYPE_REGISTRY_MANUAL.md` und `wp04_retriever_scoring.md` integrieren (ohne Informationsverlust).
|
||||
3. Ist-Stand vs. Roadmap klar kennzeichnen (WP04b–08, WP05–07).
|
||||
|
||||
Dieses Vorgehen hält die technische Wahrheit konsistent und liefert eine stabile Basis für alle nachfolgenden Dokumente.
|
||||
@@ -0,0 +1,772 @@
|
||||
# mindnet v2.2 — Programmplan
|
||||
**Version:** 2.2
|
||||
**Stand:** 2025-12-06
|
||||
**Status:** Aktiv (Fortschreibung von v2.1)
|
||||
|
||||
---
|
||||
|
||||
## 1. Programmauftrag
|
||||
|
||||
mindnet v2.2 entwickelt ein persönliches, wachsendes KI-Gedächtnis, das:
|
||||
|
||||
- Wissen, Erfahrungen, Werte und Entscheidungen speichert,
|
||||
- diese Informationen semantisch verknüpft und rekonstruierbar macht,
|
||||
- einen KI-Zwilling aufbaut, der ähnlich argumentiert, entscheidet und reflektiert wie du,
|
||||
- über mehrere Kanäle gefüttert wird:
|
||||
- Obsidian-Markdown (primäre Quelle),
|
||||
- Chat-basierter Agent (mindnet-Assistent),
|
||||
- später: Interview-Assistent (strukturierte Dialogerfassung),
|
||||
- automatisch neue Zusammenhänge erkennt und vernetzt (Edges, Typen, Hinweise),
|
||||
- sich durch Rückmeldungen (Feedback) selbst verbessert (Self-Tuning).
|
||||
|
||||
Langfristig soll mindnet als **digitales Gegenüber** funktionieren, das:
|
||||
|
||||
- dich versteht,
|
||||
- deine Denkweise reflektiert,
|
||||
- deine Werte kennt und verwendet,
|
||||
- Entscheidungen begründen kann („Warum?“),
|
||||
- Erinnerungen einordnet,
|
||||
- und für deine Nachkommen als dialogfähiger Gesprächspartner zur Verfügung steht.
|
||||
|
||||
mindnet ist **kein statisches Wissensarchiv**, sondern ein **lebendes, lernendes Gedächtnismodell** mit Fokus auf:
|
||||
|
||||
- persönliche Perspektive,
|
||||
- erklärbare Begründungspfade,
|
||||
- kontinuierliche Erweiterbarkeit,
|
||||
- robuste technische Basis (lokal, nachvollziehbar, versioniert).
|
||||
|
||||
---
|
||||
|
||||
## 2. Vision
|
||||
|
||||
> „Ein persönliches semantisches Gedächtnis, das mit mir wächst, meine Persönlichkeit spiegelt und mir als intelligenter, erklärbarer, lernender Begleiter dient.“
|
||||
|
||||
Kernprinzipien der Vision:
|
||||
|
||||
- **Erklärbarkeit:**
|
||||
Jede Antwort muss über Edges, Scores, Werte-Bezüge und Herkunftsnotizen begründbar sein.
|
||||
|
||||
- **Wachstum:**
|
||||
Das System arbeitet von Anfang an mit unvollständigen Daten, kann aber schrittweise dichter werden, ohne dass alte Notizen massenhaft manuell angepasst werden müssen.
|
||||
|
||||
- **Flexibilität (Late Binding):**
|
||||
Semantik wird überwiegend in Konfiguration (z. B. `types.yaml`, Policies, Knowledge-Design) festgelegt, nicht in jeder einzelnen Markdown-Datei.
|
||||
|
||||
- **Autonomie & Self-Healing:**
|
||||
mindnet schlägt fehlende Typen, Relationen und Edges vor (z. B. aus Inline-Relationen, Edge-Defaults, Ähnlichkeitsbeziehungen) und baut damit einen „self-healing graph“ auf.
|
||||
|
||||
- **Lernen & Self-Tuning:**
|
||||
Feedback zu Antworten (gut/schlecht, relevant/nicht relevant) fließt in Score-Gewichte, Policies und ggf. Edge-Struktur ein.
|
||||
|
||||
- **Persönlichkeit:**
|
||||
Entscheidungen werden wert- und erfahrungsbasiert begründet; mittelfristig entsteht ein Persönlichkeitsmodell, das den KI-Zwilling trägt.
|
||||
|
||||
- **Incremental Growth:**
|
||||
Das System muss mit wenigen, heterogenen Notizen starten und sich fortlaufend verdichten können – ohne Retro-Massenmigrationen im Vault.
|
||||
|
||||
---
|
||||
|
||||
## 3. Programmziele
|
||||
|
||||
### 3.1 Kurzfristig (nächste Monate)
|
||||
|
||||
- Automatische Sammlung von Wissen aus Markdown-Vault und ersten Chat-Dialogen.
|
||||
- Aufbau eines stabilen Graph-Speichers in Qdrant (`*_notes`, `*_chunks`, `*_edges`).
|
||||
- Semantischer und graphbasierter Retriever (WP-04a abgeschlossen, WP-04b/04c in Vorbereitung).
|
||||
- Automatisierte Erkennung von Beziehungen:
|
||||
- Wikilinks, Inline-Relationen, Callout-Edges, Typ-Defaults.
|
||||
- Schrittweises Lernen über Feedback (Score-Tuning, noch „manuell“ konfiguriert). :contentReference[oaicite:8]{index=8}
|
||||
- „Mitwachsendes“ Schema ohne Obsidian-Umstrukturierungen:
|
||||
- Neues Wissen kann sofort erfasst werden,
|
||||
- bestehende Notizen bleiben gültig (Virtual Schema Layer).
|
||||
- Technische Basis für die späteren Agenten und den KI-Zwilling:
|
||||
- lokaler FastAPI-Dienst,
|
||||
- Qdrant als Graph-Backend,
|
||||
- konfigurierbarer Retriever (`retriever.yaml`).
|
||||
|
||||
### 3.2 Mittelfristig
|
||||
|
||||
- Persönlichkeitsprofil wird stabil erkannt und für Antworten genutzt (Typen, Werte-Notizen, Entscheidungsnotizen).
|
||||
- Why-Layer: KI-Assistenz für Entscheidungen mit nachvollziehbarer Argumentation.
|
||||
- Interview-Assistent erstellt neue Notes automatisch (strukturierte Dialoge → Markdown).
|
||||
- mindnet erzeugt Vorschläge für neue Notes & Edges und bietet einen „Vernetzungs-Assistenten“ für manuell angelegte Notizen.
|
||||
- Agenten können über MCP-Tools (`mindnet_query`, `mindnet_subgraph`, `mindnet_store_note`, `mindnet_explain`) auf mindnet zugreifen.
|
||||
|
||||
### 3.3 Langfristig
|
||||
|
||||
- KI-Zwilling, der deinen Stil, deine Werte und deine typische Denkweise repräsentiert.
|
||||
- Weitgehend autonomes Self-Tuning der Wissensstruktur (auf Basis gespeicherter Feedback-Daten & Policies).
|
||||
- Lebenslanges Gedächtnis, das
|
||||
- konsistent bleibt,
|
||||
- wächst,
|
||||
- sich weiterentwickelt,
|
||||
- und deine Geschichte für deine Nachkommen erzählbar und erlebbar macht.
|
||||
|
||||
---
|
||||
|
||||
## 4. Architekturprinzipien
|
||||
|
||||
Die folgenden Prinzipien steuern alle Workpackages und Entscheidungen:
|
||||
|
||||
1. **Late Binding (späte Semantik)**
|
||||
- Struktur und Interpretation werden in Konfigurationen (z. B. `types.yaml`, Policies, `knowledge_design.md`) definiert, nicht direkt in den Vault-Dateien.
|
||||
- Erlaubt spätere Anpassungen ohne manuelle Massenänderungen.
|
||||
|
||||
2. **Virtual Schema Layer**
|
||||
- Typen, Regeln, Policies, Edge-Defaults werden im Schema-Layer beschrieben (z. B. `knowledge_design.md`, `mindnet_functional_architecture.md`, `types.yaml`).
|
||||
- Markdown-Dateien benötigen nur minimale, robuste Angaben (Titel, Typ, optionale Properties).
|
||||
|
||||
3. **Self-Healing Graph**
|
||||
- Der Graph wird regelmäßig analysiert (Edges, Centrality, fehlende Links).
|
||||
- Automatisierte Jobs ergänzen fehlende Kanten (ähnliche Inhalte, typisierte Standardbeziehungen, Backlinks).
|
||||
|
||||
4. **Deterministische IDs**
|
||||
- Notes, Chunks und Edges erhalten stabile IDs (z. B. UUIDv5, deterministische `note_id`/`chunk_id`/`edge_id`).
|
||||
- Der Graph ist jederzeit wiederaufbaubar (Import-Pipeline idempotent).
|
||||
|
||||
5. **Full Explainability**
|
||||
- Jeder Score, jede Beziehung, jeder Vorschlag muss über:
|
||||
- Edges,
|
||||
- Scoring-Komponenten (Semantik, Typ, Edge-Bonus, Centrality),
|
||||
- Provenienz (Notizen, Chunks, rule_id, confidence) nachvollziehbar sein.
|
||||
|
||||
6. **Persistence First**
|
||||
- Obsidian bleibt die primäre Quelle der Wahrheit.
|
||||
- mindnet ergänzt, schlägt Änderungen vor, schreibt aber nur kontrolliert (z. B. über Rewriter oder Vernetzungs-Assistent mit Freigabe).
|
||||
|
||||
7. **Minimalinvasives Schreiben**
|
||||
- Automatische Veränderungen an Markdown-Dateien erfolgen ausschließlich nach expliziter Zustimmung.
|
||||
- Rewriter und Vernetzungs-Assistent arbeiten grundsätzlich im Vorschlagsmodus.
|
||||
|
||||
8. **Incremental Growth & Early Value**
|
||||
- Das System muss bereits mit wenigen Notizen und einem kleinen Vault sinnvolle Antworten liefern (Retrieval-MVP).
|
||||
- Spätere Erweiterungen (Typen, Policies, zusätzliche Edges) dürfen bestehende Notizen nicht „ungültig“ machen, sondern ergänzen sie.
|
||||
- Feedback-Mechanismen werden früh eingeführt, auch wenn sie anfangs noch nicht vollautomatisch ausgewertet werden.
|
||||
|
||||
9. **Observability & Testbarkeit**
|
||||
- Jeder Importlauf, jede Retriever-Anfrage und jede Policy-Änderung soll prüfbar sein (Logs, Traces, Tests, Reports).
|
||||
|
||||
---
|
||||
|
||||
## 5. Programmstruktur (Phasenmodell)
|
||||
|
||||
Phase A – Fundament & Import
|
||||
Phase B – Semantik, Graph & Lernen
|
||||
Phase C – Persönlichkeitsmodell & KI-Zwilling
|
||||
Phase D – Agenten, MCP & Interaktion
|
||||
Phase E – Review, Refactoring, Dokumentation
|
||||
|
||||
Alle Workpackages sind einer Phase zugeordnet; einige Workpackages (z. B. WP-03, WP-04a) sind bereits abgeschlossen und bilden heute die Basis für alle weiteren Schritte.
|
||||
|
||||
---
|
||||
|
||||
## 6. Workpackages – detaillierte Übersicht
|
||||
|
||||
### Legende Aufwand / Komplexität
|
||||
|
||||
- Aufwand: Niedrig / Mittel / Hoch
|
||||
- Komplexität: Niedrig / Mittel / Hoch
|
||||
|
||||
Die Status-Ampel zeigt den aktuellen Stand gemäß Programmfortschritt. :contentReference[oaicite:18]{index=18}
|
||||
|
||||
---
|
||||
|
||||
### WP-01 – Wissensdesign (abgeschlossen)
|
||||
|
||||
**Phase:** A
|
||||
**Status:** 🟢 abgeschlossen
|
||||
|
||||
**Ziel:**
|
||||
Grundlegendes Schema für mindnet festlegen (`knowledge_design.md`), inkl.:
|
||||
|
||||
- Note-Typen (z. B. `concept`, `experience`, `project`, `decision`, `value`, `principle`, `person`, `event`, etc.).
|
||||
- Standard-Properties (Titel, Typ, Tags, Status, Zeitstempel, Referenzen, Policies).
|
||||
- Grundstruktur von Edges (z. B. `belongs_to`, `references`, `depends_on`, `similar_to`, `related_to`, `influenced_by`).
|
||||
- Abbildung zentraler Lebens-Bereiche, Rollen und Wissenskategorien.
|
||||
|
||||
**Aufwand / Komplexität:**
|
||||
- Aufwand: Mittel
|
||||
- Komplexität: Mittel
|
||||
|
||||
**Erreichte Ergebnisse / Funktionsumfang:**
|
||||
|
||||
- `knowledge_design.md` in Version 1.5 definiert:
|
||||
- Zielarchitektur,
|
||||
- Collections,
|
||||
- Identitätsmodell,
|
||||
- zentrale Typen und Edges. :contentReference[oaicite:20]{index=20}
|
||||
- Das Dokument beschreibt klar:
|
||||
- wie Wissen fachlich strukturiert wird,
|
||||
- welche Typen und Relationen in mindnet „erwartet“ werden,
|
||||
- wie spätere Erweiterungen (neue Typen, zusätzliche Edge-Arten) ohne Retro-Migration möglich sind.
|
||||
|
||||
**DoD:**
|
||||
|
||||
- Schema ist konsistent beschrieben.
|
||||
- Grundtypen und Edge-Arten sind definiert und im Code konzeptionell verankert (z. B. via `types.yaml`).
|
||||
- Das Dokument dient als Referenz für alle folgenden WPs.
|
||||
|
||||
---
|
||||
|
||||
### WP-02 – Chunking & Hash-Strategie (abgeschlossen)
|
||||
|
||||
**Phase:** A
|
||||
**Status:** 🟢 abgeschlossen
|
||||
|
||||
**Ziel:**
|
||||
Festlegen und implementieren einer robusten Chunking-Strategie und einer Hash-Logik zur Änderungserkennung.
|
||||
|
||||
**Umfang (fachlich & technisch):**
|
||||
|
||||
- Text in semantisch sinnvolle Chunks teilen (Heading-/Block-aware).
|
||||
- Parallelhaltung von:
|
||||
- `text` (Chunk-Kern ohne Overlap),
|
||||
- `window` (Chunk-Fenster mit Overlap für Embeddings).
|
||||
- Hash-Modi zur Änderungserkennung (Body / Frontmatter / Full, canonical vs. none).
|
||||
- Unterstützung eines „Baseline-Modus“ (mehrere Hash-Signaturen), um spätere Modus-Wechsel zu ermöglichen, ohne alle Notizen als „geändert“ zu sehen.
|
||||
|
||||
**Aufwand / Komplexität:**
|
||||
- Aufwand: Mittel
|
||||
- Komplexität: Mittel/Hoch
|
||||
|
||||
**Erreichte Ergebnisse / Funktionsumfang:**
|
||||
|
||||
- Implementierte Chunker-Logik mit Ziel-Längen, Overlap und Nachbarschaft (`neighbors_prev`, `neighbors_next`).
|
||||
- Tests zur Überprüfung der Rekonstruktion von Notiz-Bodies aus Chunks (`verify_chunks_integrity.py`, `check_chunks_window_vs_text.py`).
|
||||
- Hash-Konfiguration über ENV-Variablen + Option `--baseline-modes`.
|
||||
- Roundtrip-Tests für Vault → Qdrant → Export-Vault (keine ungewollten Veränderungen an Inhalten).
|
||||
|
||||
**DoD:**
|
||||
|
||||
- Chunking stabil, deterministisch, überprüft.
|
||||
- Hash-Strategie ermöglicht sichere Änderungserkennung ohne Mass-Diffs bei Konfigurationswechsel.
|
||||
|
||||
---
|
||||
|
||||
### WP-03 – Import-Pipeline & Edge-System v2 (abgeschlossen)
|
||||
|
||||
**Phase:** A/B
|
||||
**Status:** 🟢 abgeschlossen
|
||||
|
||||
**Ziel:**
|
||||
Durchgängige Import-Pipeline von Markdown → Notes/Chunks/Edges in Qdrant mit deterministischen IDs, typbasierten Defaults und erklärbarem Edge-Modell.
|
||||
|
||||
**Umfang:**
|
||||
|
||||
- Parser (robust gegenüber Zeichensatzfehlern, Frontmatter/Body-Erkennung, Sonderfälle wie Silverbullet-Dateien).
|
||||
- Bau von Note-Payloads (`note_payload`):
|
||||
- Typ-Vererbung (`retriever_weight`, `chunk_profile`, `edge_defaults`).
|
||||
- Bau von Chunk-Payloads (`chunk_payload`):
|
||||
- `text`, `window`, Nachbarn, Typ-Information.
|
||||
- Edge-System:
|
||||
- Strukturkanten (`belongs_to`, `next`, `prev`),
|
||||
- explizite Wikilinks (`[[Note]]` → `references`),
|
||||
- Inline-Relationen (`[[rel:depends_on Note]]`),
|
||||
- Callout-Edges (`> [!edge] related_to: [[Note]]`),
|
||||
- typbasierte Standardkanten aus `types.yaml` (`edge_defaults`).
|
||||
- Idempotente Upserts in Qdrant (Notes/Chunks/Edges).
|
||||
|
||||
**Aufwand / Komplexität:**
|
||||
- Aufwand: Hoch
|
||||
- Komplexität: Hoch
|
||||
|
||||
**Erreichte Ergebnisse / Funktionsumfang:**
|
||||
|
||||
- Vollständige E2E-Import-Pipeline incl. Tests (`run_e2e_roundtrip.sh`, `compare_vaults.py`, `validate_edges`, `edges_full_check`).
|
||||
- deterministische ID-Vergabe für Notes/Chunks/Edges.
|
||||
- vollständiges Edge-Modell gemäß Facharchitektur (`mindnet_functional_architecture.md`). :contentReference[oaicite:26]{index=26}
|
||||
- Exporte zurück nach Markdown (Export-Pipeline) ermöglicht Roundtrip-Vergleich.
|
||||
|
||||
**DoD:**
|
||||
|
||||
- Importer generiert stabile Notes/Chunks/Edges für gesamten Vault.
|
||||
- Edge-Invarianten (`belongs_to == #Chunks`, `next/prev` konsistent) erfüllt.
|
||||
- Roundtrip: Exportierte Vault-Version stimmt inhaltlich mit originalem Vault überein (bis auf bewusst ausgeschlossene Dateien).
|
||||
|
||||
---
|
||||
|
||||
### WP-04a – Retriever & Graph-Scoring (abgeschlossen)
|
||||
|
||||
**Phase:** B
|
||||
**Status:** 🟢 abgeschlossen
|
||||
|
||||
**Ziel:**
|
||||
Aufbau eines hybriden Retrievers, der Semantik (Embeddings), Typwissen (`retriever_weight`) und Graph-Informationen (Edges, Centrality) zu einem erklärbaren Score kombiniert.
|
||||
|
||||
**Umfang:**
|
||||
|
||||
- Retriever-Modul (`app/core/retriever.py`):
|
||||
- Vektorsuche über `*_chunks`,
|
||||
- Zusammenführung von Chunk-Treffern mit Note-Metadaten,
|
||||
- Aufbau eines Subgraphen (Nachbarn über Edges, Centrality-Berechnung).
|
||||
- Scoring-Formel:
|
||||
semantic_score
|
||||
retriever_weight
|
||||
edge_bonus
|
||||
centrality_bonus
|
||||
- Konfiguration über `config/retriever.yaml` (Gewichte, Parameter). :contentReference[oaicite:28]{index=28}
|
||||
- FastAPI-Endpoint `/query` mit Rückgabe:
|
||||
- Trefferliste,
|
||||
- Teil-Scores,
|
||||
- Begründung (inkl. Edge-Kontext).
|
||||
|
||||
**Aufwand / Komplexität:**
|
||||
- Aufwand: Hoch
|
||||
- Komplexität: Hoch
|
||||
|
||||
**Erreichte Ergebnisse / Funktionsumfang:**
|
||||
|
||||
- Hybrides Scoring-Modell mit konfigurierbaren Gewichten (`semantic_weight`, `edge_weight`, `centrality_weight`).
|
||||
- Nutzung von `retriever_weight` aus `types.yaml` in Score-Berechnung.
|
||||
- Graph-boni aus echten Edges (explizit, Inline, Typ-Defaults, Strukturkanten).
|
||||
- Funktionsfähiger `/query`-Endpoint mit Testcases (inkl. Relations-Showcase-Vault).
|
||||
- Basis für spätere Self-Tuning-Mechanismen (Logging von Scores, Konfiguration in `retriever.yaml`).
|
||||
|
||||
**DoD:**
|
||||
|
||||
- `/query` liefert nachvollziehbare Ergebnisse mit sichtbarem Einfluss von:
|
||||
- Typprioritäten,
|
||||
- Edge-Struktur,
|
||||
- Centrality.
|
||||
- Scoring ist rein konfigurierbar (Konfig-Change → Score-Veränderung ohne Code-Änderung).
|
||||
|
||||
---
|
||||
|
||||
### WP-04b – Explanation Layer (geplant)
|
||||
|
||||
**Phase:** B
|
||||
**Status:** 🟡 geplant
|
||||
|
||||
**Ziel:**
|
||||
Treffererklärungen liefern, die für dich (Mindmaster) und spätere Nutzer verständlich sind (Why-Layer).
|
||||
|
||||
**Umfang:**
|
||||
|
||||
- Erklärungsschicht über dem Retriever:
|
||||
- Aufschlüsselung des Scores pro Treffer,
|
||||
- Ketten von Edges (Pfade) als Begründungsgraph,
|
||||
- Markierung relevanter Chunks / Notizen.
|
||||
- Darstellung von:
|
||||
- verwendeten Typen,
|
||||
- Edges und deren `rule_id` / `confidence`,
|
||||
- Centrality-Beiträgen.
|
||||
|
||||
**Aufwand / Komplexität:**
|
||||
- Aufwand: Mittel
|
||||
- Komplexität: Hoch
|
||||
|
||||
**DoD (Zielbild):**
|
||||
|
||||
- API-Endpoint (z. B. `/query/explain`), der eine strukturierte Erklärung in JSON liefert.
|
||||
- Mindestens ein Beispiel, in dem Score-Änderung nachvollziehbar auf Policies/Edges zurückgeführt werden kann.
|
||||
|
||||
---
|
||||
|
||||
### WP-04c – Feedback Logging & Bewertungsdaten (geplant)
|
||||
|
||||
**Phase:** B
|
||||
**Status:** 🟡 geplant
|
||||
|
||||
**Ziel:**
|
||||
Grundlage für Self-Tuning schaffen, indem Queries, Antworten und Bewertungen systematisch protokolliert werden.
|
||||
|
||||
**Umfang:**
|
||||
|
||||
- Logging von:
|
||||
- Query-Texten,
|
||||
- verwendeten Policies/Config-Versionen,
|
||||
- Trefferlisten (Scores, Edges),
|
||||
- Feedback (z. B. „hilfreich“, „irrelevant“, „falsch“).
|
||||
- Persistente Speicherung (z. B. in einer separaten Sammlung oder Datei-basiert).
|
||||
- Bereitstellung von Auswertungs-Skripten (Statistiken, Ranking-Qualität).
|
||||
|
||||
**Aufwand / Komplexität:**
|
||||
- Aufwand: Mittel
|
||||
- Komplexität: Mittel/Hoch
|
||||
|
||||
**DoD (Zielbild):**
|
||||
|
||||
- Vollständiger Feedback-Trail für ausgewählte Test-Sessions.
|
||||
- Exportfähige Datenbasis zur späteren Optimierung der Retriever-Parameter.
|
||||
|
||||
---
|
||||
|
||||
### WP-05 – Persönlichkeitsmodell (geplant)
|
||||
|
||||
**Phase:** C
|
||||
**Status:** 🟡 geplant
|
||||
|
||||
**Ziel:**
|
||||
Aufbau eines expliziten Persönlichkeitsmodells, das Werte, Prinzipien, prägende Erfahrungen und wiederkehrende Entscheidungsstrategien abbildet.
|
||||
|
||||
**Umfang:**
|
||||
|
||||
- Identifikation und Typisierung relevanter Notes:
|
||||
- `value`, `principle`, `life_event`, `milestone`, `decision`.
|
||||
- Aufbau eines Modells, das:
|
||||
- typische Begründungsmuster erkennt,
|
||||
- Wertprioritäten identifiziert,
|
||||
- langfristig in Scoring & Antwort-Formulierung einfließt.
|
||||
|
||||
**Aufwand / Komplexität:**
|
||||
- Aufwand: Mittel
|
||||
- Komplexität: Mittel
|
||||
|
||||
**DoD (Zielbild):**
|
||||
|
||||
- Dokumentiertes Persönlichkeitsmodell (z. B. YAML/Markdown), das konkrete Wert-/Prinzipien-Strukturen beschreibt.
|
||||
- Basis-Integration in den Retriever (z. B. Wert-Bezüge in Explanation-Layer).
|
||||
|
||||
---
|
||||
|
||||
### WP-06 – Decision Engine (geplant)
|
||||
|
||||
**Phase:** C
|
||||
**Status:** 🟡 geplant
|
||||
|
||||
**Ziel:**
|
||||
Entscheidungsunterstützung auf Basis von Wissen, Persönlichkeit und Zielen – inklusive nachvollziehbarer Begründung.
|
||||
|
||||
**Umfang:**
|
||||
|
||||
- Identifikation typischer Entscheidungssituationen (z. B. beruflich, familiär, gesundheitlich).
|
||||
- Integration von:
|
||||
- fachlichem Wissen,
|
||||
- Erfahrungen (Vergangenheitsentscheidungen),
|
||||
- Persönlichkeitsmodell.
|
||||
- Ableitung von Entscheidungs-Templates / Heuristiken.
|
||||
|
||||
**Aufwand / Komplexität:**
|
||||
- Aufwand: Mittel
|
||||
- Komplexität: Hoch
|
||||
|
||||
**DoD (Zielbild):**
|
||||
|
||||
- Prototypischer „Decision-Assistant“, der für definierte Szenarien begründete Empfehlungen gibt.
|
||||
- Erklärungen referenzieren konkrete Notes/Edges.
|
||||
|
||||
---
|
||||
|
||||
### WP-07 – Interview-Assistent (geplant)
|
||||
|
||||
**Phase:** C
|
||||
**Status:** 🟡 geplant
|
||||
|
||||
**Ziel:**
|
||||
Dialogbasierter Erfassungs-Assistent, der strukturierte Interviews führt und daraus konsistente Markdown-Notizen generiert.
|
||||
|
||||
**Umfang:**
|
||||
|
||||
- Design von Interview-Flows (z. B. für Lebensereignisse, Projekte, Werte-Workshops).
|
||||
- Konvertierung von Dialogtranskripten in typisierte Notes (inkl. Typ, Tags, erste Edges).
|
||||
- Integration mit Vault-Struktur (Ablagepfade, Namenskonventionen).
|
||||
|
||||
**Aufwand / Komplexität:**
|
||||
- Aufwand: Niedrig/Mittel
|
||||
- Komplexität: Mittel
|
||||
|
||||
**DoD (Zielbild):**
|
||||
|
||||
- Mindestens ein produktionsnaher Flow (z. B. „Erzähle mir von einem prägenden Ereignis“).
|
||||
- Automatische Erzeugung von Markdown-Notizen, die vom Importer direkt verarbeitet werden können.
|
||||
|
||||
---
|
||||
|
||||
### WP-08 – Self-Tuning v1/v2 (geplant)
|
||||
|
||||
**Phase:** B/C
|
||||
**Status:** 🟡 geplant
|
||||
|
||||
**Ziel:**
|
||||
Aufbau eines Self-Tuning-Mechanismus, der auf Basis von Feedback-Daten Vorschläge für Retriever- und Policy-Anpassungen macht.
|
||||
|
||||
**Umfang:**
|
||||
|
||||
- Auswertung der Feedback-Daten (WP-04c).
|
||||
- Regel-basierte Anpassungs-Vorschläge:
|
||||
- Gewichte in `retriever.yaml`,
|
||||
- Typ-Prioritäten,
|
||||
- Edge-Provenance-Gewichte.
|
||||
- Optionale halbautomatische Übernahme (Freigabe durch Mindmaster).
|
||||
|
||||
**Aufwand / Komplexität:**
|
||||
- Aufwand: Hoch
|
||||
- Komplexität: Hoch
|
||||
|
||||
**DoD (Zielbild):**
|
||||
|
||||
- Mindestens ein in der Praxis getesteter Tuning-Zyklus:
|
||||
- Feedback sammeln,
|
||||
- Vorschlag generieren,
|
||||
- Anpassung vornehmen,
|
||||
- Qualitätsgewinn nachweisen.
|
||||
|
||||
---
|
||||
|
||||
### WP-09 – Vault-Onboarding & Migration (geplant)
|
||||
|
||||
**Phase:** B
|
||||
**Status:** 🟡 geplant
|
||||
|
||||
**Ziel:**
|
||||
Sicherstellen, dass bestehende und neue Obsidian-Vaults schrittweise in mindnet integriert werden können – ohne Massenumbau.
|
||||
|
||||
**Umfang:**
|
||||
|
||||
- Tools/Guides für:
|
||||
- Analyse des Vault-Status (Typen, Edges, Lücken),
|
||||
- Empfehlungen für minimale Anpassungen (z. B. Typ-Setzung, Basis-Tags),
|
||||
- optional automatisierte Ergänzung von Properties (mit Freigabe).
|
||||
- Unterstützung eines „wachsenden Systems“:
|
||||
- erste Notizen liefern sofort Mehrwert im Retriever,
|
||||
- später können Typen und Edges ergänzt werden, ohne alte Notizen zu brechen.
|
||||
|
||||
**Aufwand / Komplexität:**
|
||||
- Aufwand: Mittel
|
||||
- Komplexität: Niedrig/Mittel
|
||||
|
||||
**DoD (Zielbild):**
|
||||
|
||||
- Dokumentierter Onboarding-Prozess für Produkt-Vault.
|
||||
- Mindestens ein erfolgreicher Durchlauf mit deinem realen Vault (Pilot).
|
||||
|
||||
---
|
||||
|
||||
### WP-10 – Chat-Interface & Writeback (geplant)
|
||||
|
||||
**Phase:** D
|
||||
**Status:** 🟡 geplant
|
||||
|
||||
**Ziel:**
|
||||
Chat-basierter mindnet-Agent, der sowohl Fragen beantwortet als auch neue Notizen erzeugen/aktualisieren kann (Writeback Richtung Vault).
|
||||
|
||||
**Umfang:**
|
||||
|
||||
- Chat-Frontend (z. B. in deiner bestehenden LLM-Umgebung).
|
||||
- Anbindung an:
|
||||
- `/query` (Retriever),
|
||||
- MCP-Tools (WP-13),
|
||||
- Import-/Export-Routinen.
|
||||
- Funktionen:
|
||||
- Antworten aus mindnet-Wissen,
|
||||
- Vorschlag neuer Notes/Edges,
|
||||
- optionales Schreiben ins Vault (Markdown-Erzeugung, Rewriting nur mit Freigabe).
|
||||
|
||||
**Aufwand / Komplexität:**
|
||||
- Aufwand: Hoch
|
||||
- Komplexität: Hoch
|
||||
|
||||
**DoD (Zielbild):**
|
||||
|
||||
- Du kannst im Chat:
|
||||
- Fragen an dein Wissensnetz stellen,
|
||||
- neue Gedanken/Erinnerungen erfassen,
|
||||
- Vorschläge für Ablage & Vernetzung als Markdown-Notizen erhalten.
|
||||
|
||||
---
|
||||
|
||||
### WP-11 – Knowledge-Builder & Vernetzungs-Assistent (geplant)
|
||||
|
||||
**Phase:** D
|
||||
**Status:** 🟡 geplant
|
||||
|
||||
**Ziel:**
|
||||
Assistent, der manuell erstellte oder importierte Notizen analysiert und Vorschläge für Typen, Edges und Einordnung macht.
|
||||
|
||||
**Umfang:**
|
||||
|
||||
- Analyse neuer/aktualisierter Notes:
|
||||
- mögliche Typen,
|
||||
- passende Tags,
|
||||
- potenzielle Relationen zu bestehenden Notes.
|
||||
- Vorschläge:
|
||||
- Inline-Relationen,
|
||||
- Wikilinks,
|
||||
- Typ-Zuweisung.
|
||||
- Optional: Generierung konkreter Patch-Vorschläge für Markdown-Dateien (Diff-basiert).
|
||||
|
||||
**Aufwand / Komplexität:**
|
||||
- Aufwand: Hoch
|
||||
- Komplexität: Hoch
|
||||
|
||||
**DoD (Zielbild):**
|
||||
|
||||
- Mindestens ein produktionsnaher Workflow, bei dem ein neuer Eintrag:
|
||||
- erstellt wird,
|
||||
- analysiert wird,
|
||||
- Vernetzungs-Vorschläge liefert,
|
||||
- und nach Freigabe in den Vault zurückgeschrieben wird.
|
||||
|
||||
---
|
||||
|
||||
### WP-12 – Knowledge Rewriter (Soft Mode, geplant)
|
||||
|
||||
**Phase:** C/D
|
||||
**Status:** 🟡 geplant
|
||||
|
||||
**Ziel:**
|
||||
Optionaler Assistent, der vorhandene Notes „aufräumt“, zusammenfasst oder reorganisiert – ausschließlich mit expliziter Freigabe. :contentReference[oaicite:31]{index=31}
|
||||
|
||||
**Einschränkungen:**
|
||||
|
||||
- Änderungen an Markdown-Dateien nur nach Freigabe.
|
||||
- Kein „silent overwrite“ existierender Inhalte.
|
||||
|
||||
**Aufwand / Komplexität:**
|
||||
- Aufwand: Niedrig/Mittel
|
||||
- Komplexität: Mittel
|
||||
|
||||
**DoD (Zielbild):**
|
||||
|
||||
- Rewriter generiert Vorschläge (neue Versionen) inklusive Diffs.
|
||||
- Nach Freigabe werden Notes konsistent aktualisiert:
|
||||
- IDs, Edges und Typen bleiben konsistent oder werden sauber mitaktualisiert.
|
||||
|
||||
---
|
||||
|
||||
### WP-13 – MCP-Integration & Agenten-Layer (geplant)
|
||||
|
||||
**Phase:** D
|
||||
**Status:** 🟡 geplant
|
||||
|
||||
**Ziel:**
|
||||
mindnet als MCP-Server bereitstellen, damit Agenten/LLMs standardisierte Tools nutzen können (z. B. in deiner n8n- und LLM-Umgebung).
|
||||
|
||||
**Umfang:**
|
||||
|
||||
- MCP-Server mit Tools wie:
|
||||
- `mindnet_query`,
|
||||
- `mindnet_subgraph`,
|
||||
- `mindnet_store_note`,
|
||||
- `mindnet_explain`.
|
||||
- Sichere Konfiguration (lokal, authentifiziert, logging-fähig).
|
||||
|
||||
**Aufwand / Komplexität:**
|
||||
- Aufwand: Mittel
|
||||
- Komplexität: Mittel
|
||||
|
||||
**DoD (Zielbild):**
|
||||
|
||||
- Lokaler MCP-Server läuft und ist aus einem LLM-Client nutzbar.
|
||||
- Tools liefern konsistente Ergebnisse (aligned mit FastAPI-Endpoints).
|
||||
- Sicherheits- und Loggingkonzept dokumentiert.
|
||||
|
||||
---
|
||||
|
||||
### WP-14 – Review / Refactoring / Dokumentation (geplant)
|
||||
|
||||
**Phase:** E
|
||||
**Status:** 🟡 geplant
|
||||
|
||||
**Ziel:**
|
||||
Aufräumen, dokumentieren, stabilisieren – insbesondere für Onboarding Dritter und Langfrist-Betrieb.
|
||||
|
||||
**Deliverables (Beispiele):**
|
||||
|
||||
- konsolidierte technische Architektur (`mindnet_technical_architecture.md`),
|
||||
- fachliche Architektur (`mindnet_functional_architecture.md`),
|
||||
- Identity-Modell (`identity_model.md`),
|
||||
- API-Spezifikation (`api_spec.md`),
|
||||
- MCP-Integration (`mcp_integration.md`),
|
||||
- aktualisiertes Handbuch (`Handbuch.md`).
|
||||
|
||||
**Aufwand / Komplexität:**
|
||||
- Aufwand: Mittel
|
||||
- Komplexität: Niedrig/Mittel
|
||||
|
||||
**DoD (Zielbild):**
|
||||
|
||||
- Aktueller Stand ist sauber dokumentiert.
|
||||
- Wichtige Designentscheidungen sind nachvollziehbar festgehalten.
|
||||
- Onboarding einer dritten Person (z. B. Entwickler:in) ist mit Dokumentation möglich.
|
||||
|
||||
---
|
||||
|
||||
## 7. Abhängigkeiten (vereinfacht, aktualisiert)
|
||||
|
||||
WP01 → WP02 → WP03 → WP04a
|
||||
WP04a → WP04b → WP04c → WP08
|
||||
WP03 → WP05 → WP06 → WP07
|
||||
WP03 → WP09
|
||||
WP01/WP03 → WP10 → WP11 → WP12
|
||||
WP03/WP04 → WP13
|
||||
Alles → WP14
|
||||
|
||||
Diese Struktur bleibt gegenüber v2.1 unverändert, ist aber jetzt explizit an den bereits implementierten Stand (WP-03, WP-04a) angepasst.
|
||||
|
||||
---
|
||||
|
||||
## 8. Laufzeit- & Komplexitätsindikatoren (unverändert, ergänzt)
|
||||
|
||||
| WP | Aufwand | Komplexität |
|
||||
|-------|----------------|------------------|
|
||||
| WP04b | Mittel | Hoch |
|
||||
| WP04c | Mittel | Mittel/Hoch |
|
||||
| WP05 | Mittel | Mittel |
|
||||
| WP06 | Mittel | Hoch |
|
||||
| WP07 | Niedrig/Mittel | Mittel |
|
||||
| WP08 | Hoch | Hoch |
|
||||
| WP09 | Mittel | Niedrig/Mittel |
|
||||
| WP10 | Hoch | Hoch |
|
||||
| WP11 | Hoch | Hoch |
|
||||
| WP12 | Niedrig/Mittel | Mittel |
|
||||
| WP13 | Mittel | Mittel |
|
||||
| WP14 | Mittel | Niedrig/Mittel | :contentReference[oaicite:35]{index=35}
|
||||
|
||||
Für WP01–WP04a gilt: Aufwand/Komplexität wie in v2.1; Status jetzt explizit „abgeschlossen“ mit beschriebenem Funktionsumfang.
|
||||
|
||||
---
|
||||
|
||||
## 9. Programmfortschritt (Ampel, aktualisiert)
|
||||
|
||||
| WP | Status |
|
||||
|-------|--------|
|
||||
| WP01 | 🟢 |
|
||||
| WP02 | 🟢 |
|
||||
| WP03 | 🟢 |
|
||||
| WP04a | 🟢 |
|
||||
| WP04b | 🟡 |
|
||||
| WP04c | 🟡 |
|
||||
| WP05 | 🟡 |
|
||||
| WP06 | 🟡 |
|
||||
| WP07 | 🟡 |
|
||||
| WP08 | 🟡 |
|
||||
| WP09 | 🟡 |
|
||||
| WP10 | 🟡 |
|
||||
| WP11 | 🟡 |
|
||||
| WP12 | 🟡 |
|
||||
| WP13 | 🟡 |
|
||||
| WP14 | 🟡 | :contentReference[oaicite:36]{index=36}
|
||||
|
||||
---
|
||||
|
||||
## 10. Governance & Versionierung
|
||||
|
||||
- Versionierung über Gitea (Branches, Tags, Releases).
|
||||
- Jede wesentliche Änderung an Architektur/Schema:
|
||||
- eigener Commit,
|
||||
- kurze Doku (z. B. in `CHANGELOG.md` oder Projekt-Notiz).
|
||||
- Jedes Workpackage erhält ein eigenes Chat-Fenster mit dediziertem Prompt (WP-Hand-Over).
|
||||
- Programmleitung verantwortet:
|
||||
- Konsistenz von Fach- und Technikdokumenten,
|
||||
- Abhängigkeiten und Priorisierung,
|
||||
- Freigabe von Meilensteinen,
|
||||
- Sicherstellung, dass Änderungen mit dem langfristigen Ziel (KI-Zwilling) vereinbar sind.
|
||||
|
||||
---
|
||||
|
||||
## 11. Executive Summary (unverändert, präzisiert)
|
||||
|
||||
mindnet v2.2 ist so aufgesetzt, dass:
|
||||
|
||||
- du **schrittweise** Wissen erfassen kannst (Obsidian + Chat + später Interview-Assistent),
|
||||
- die Struktur **mitwächst**, ohne Retro-Massenarbeit im Vault,
|
||||
- ein **hybrider Retriever** qualitativ hochwertige, erklärbare Antworten liefert,
|
||||
- ein **Self-Healing- und Self-Tuning-Mechanismus** vorbereitet ist, der auf Feedback aufsetzt,
|
||||
- ein **Persönlichkeitsmodell** und eine **Decision Engine** entstehen,
|
||||
- langfristig ein **KI-Zwilling** aufgebaut wird, der deine Werte, Erfahrungen und Denkweise spiegelt,
|
||||
- die technische Architektur (FastAPI, Qdrant, YAML-Policies, MCP-Integration) lokal, nachvollziehbar und erweiterbar bleibt.
|
||||
|
||||
Dieser Programmplan bildet die konsolidierte Grundlage (v2.2) für alle weiteren Arbeiten.
|
||||
Er integriert die bisherigen Ergebnisse (WP-01 bis WP-04a) und die zusätzlichen Anforderungen aus der späteren Planung (Feedback-basiertes Tuning, wachsendes System, Chat-/Interview-Eingabe, Vault-Writeback) in ein einheitliches, versionierbares Dokument.
|
||||
|
||||
@@ -0,0 +1,307 @@
|
||||
# mindnet v2 — Dokumentationsreview (vollständig & konsolidiert)
|
||||
|
||||
Dieses Dokument fasst alle notwendigen Anpassungen der mindnet-v2 Dokumentation zusammen.
|
||||
Der Fokus liegt ausschließlich auf Aspekten, die durch die aktuellen Arbeiten verändert wurden:
|
||||
|
||||
- neue Edge-Mechanik
|
||||
- neuer Importer
|
||||
- neues Verhalten von note_payload & chunk_payload
|
||||
- Default-Typregeln aus `types.yaml`
|
||||
- Inline- und Callout-Relations
|
||||
- WP04a – Einführung der Retriever-Schicht
|
||||
- neue Diagnose- und Testskripte
|
||||
|
||||
Es handelt sich um eine strukturelle Analyse.
|
||||
Nichts davon überschreibt bestehende Inhalte — es markiert nur, **was angepasst werden muss**.
|
||||
|
||||
---
|
||||
|
||||
# 1) knowledge_design.md
|
||||
## Erforderliche Anpassungen
|
||||
|
||||
### 1.1 Neue Edge-Arten dokumentieren
|
||||
Es müssen vier neue Kategorien beschrieben werden:
|
||||
|
||||
1. **Inline-Edges**
|
||||
|
||||
Beispiel:
|
||||
`[ [rel:depends_on Embeddings 101] ]`
|
||||
(Leerzeichen hier zur Demonstration; im echten Dokument ohne Leerzeichen)
|
||||
|
||||
2. **Callout-Edges**
|
||||
|
||||
> [!edge] related_to: [[Vector DB Basics]] [[Embeddings 101]]
|
||||
|
||||
3. **Typbasierte Default-Kanten**
|
||||
Aus `types.yaml` → `edge_defaults`
|
||||
|
||||
4. **Strukturkanten**
|
||||
- belongs_to
|
||||
- next
|
||||
- prev
|
||||
|
||||
### 1.2 Neue Edge-Felder ergänzen
|
||||
Jeder Edge enthält jetzt:
|
||||
|
||||
- `rule_id`
|
||||
- `confidence`
|
||||
- `scope = "chunk"`
|
||||
- deterministischer `edge_id`
|
||||
|
||||
Diese Felder müssen vollständig erklärt werden.
|
||||
|
||||
### 1.3 Chunk-basierte Kanten
|
||||
Neu: **Alle Kanten entstehen ausschließlich zwischen Chunks**, nie zwischen Notes.
|
||||
Notes dienen nur noch als Metadatencontainer.
|
||||
|
||||
---
|
||||
|
||||
# 2) TYPE_REGISTRY_MANUAL.md
|
||||
## Erforderliche Anpassungen
|
||||
|
||||
### 2.1 Typauflösung dokumentieren
|
||||
Die Regeln sind jetzt:
|
||||
|
||||
- Typ aus Frontmatter → höchste Priorität
|
||||
- Falls nicht gesetzt: Typ aus Dateipfad ableiten
|
||||
- Falls unbekannt: Fallback auf `types.default`
|
||||
- Typ definiert:
|
||||
- retriever_weight
|
||||
- chunk_profile
|
||||
- edge_defaults
|
||||
|
||||
Diese müssen klar dokumentiert werden.
|
||||
|
||||
### 2.2 Neue Fallback-Regeln
|
||||
Wenn Werte fehlen, werden sie ersetzt durch:
|
||||
|
||||
types.default.retriever_weight
|
||||
types.default.chunk_profile
|
||||
types.default.edge_defaults
|
||||
|
||||
### 2.3 Beispiele aktualisieren
|
||||
Alle Beispiele müssen auf die neue Inline-Syntax angepasst werden.
|
||||
|
||||
---
|
||||
|
||||
# 3) wp04_retriever_scoring.md
|
||||
## Erforderliche Anpassungen
|
||||
|
||||
### 3.1 Neue Scoring-Formel
|
||||
total_score =
|
||||
semantic_weight * semantic_score
|
||||
+ edge_weight * edge_bonus
|
||||
+ centrality_weight * centrality_bonus
|
||||
+ type_weight * retriever_weight
|
||||
|
||||
### 3.2 Bedeutung von `confidence`
|
||||
confidence wirkt multiplikativ bei edge_bonus.
|
||||
Inline > Callout > Default (Standardwerte im Code hinterlegt).
|
||||
|
||||
### 3.3 Pfadbegründungen
|
||||
Der Retriever gibt optional `paths[]` zurück:
|
||||
|
||||
paths: [
|
||||
{
|
||||
"via": "depends_on",
|
||||
"confidence": 0.7,
|
||||
"hops": ["Note A", "Chunk A#01", "Note B"]
|
||||
}
|
||||
]
|
||||
|
||||
Diese Funktion müssen dokumentiert werden.
|
||||
|
||||
---
|
||||
|
||||
# 4) mindnet_v2_implementation_playbook.md
|
||||
## Erforderliche Anpassungen
|
||||
|
||||
### 4.1 Neuer Import-Prozess vollständig dokumentieren
|
||||
Der Importer läuft jetzt so:
|
||||
|
||||
1. Markdown lesen
|
||||
2. Frontmatter extrahieren
|
||||
3. Typauflösung via `types.yaml`
|
||||
4. Note-Payload generieren
|
||||
5. Chunking anwenden
|
||||
6. Inline-Kanten finden
|
||||
7. Callout-Kanten finden
|
||||
8. Default-Edges erzeugen
|
||||
9. Strukturkanten erzeugen
|
||||
10. Chunks in Qdrant upserten
|
||||
11. Edges upserten
|
||||
12. Diagnose laufen lassen:
|
||||
|
||||
python3 -m scripts.payload_dryrun
|
||||
python3 -m scripts.edges_full_check
|
||||
|
||||
### 4.2 Erweiterte Testpipeline
|
||||
Neu zu dokumentieren:
|
||||
|
||||
tests.test_edges_all
|
||||
scripts.edges_full_check
|
||||
scripts.payload_dryrun
|
||||
|
||||
Diese Testschritte sind jetzt Standardanforderung nach jeder Änderung.
|
||||
|
||||
---
|
||||
|
||||
# 5) mindnet_technical_architecture.md
|
||||
## Erforderliche Anpassungen
|
||||
|
||||
### 5.1 Aktualisiertes Schema für Qdrant Collections
|
||||
|
||||
#### Notes:
|
||||
note_id
|
||||
title
|
||||
type
|
||||
fulltext
|
||||
retriever_weight
|
||||
chunk_profile
|
||||
edge_defaults
|
||||
tags
|
||||
updated
|
||||
|
||||
#### Chunks:
|
||||
chunk_id
|
||||
note_id
|
||||
text
|
||||
window
|
||||
ord
|
||||
retriever_weight
|
||||
chunk_profile
|
||||
neighbors_prev
|
||||
neighbors_next
|
||||
|
||||
#### Edges:
|
||||
edge_id
|
||||
source_id
|
||||
target_id
|
||||
kind
|
||||
scope = "chunk"
|
||||
rule_id
|
||||
confidence
|
||||
|
||||
### 5.2 Dokumentation der Edge-Pipeline
|
||||
Neu beschreiben:
|
||||
|
||||
- Inline-Parser
|
||||
- Callout-Parser
|
||||
- Default-Edge-Resolver
|
||||
- Strukturkanten
|
||||
- Normalisierung für Qdrant
|
||||
- deterministische ID-Bildung
|
||||
|
||||
---
|
||||
|
||||
# 6) mindnet_functional_architecture.md
|
||||
## Erforderliche Anpassungen
|
||||
|
||||
### 6.1 Inline-Relations – einzig gültige Syntax
|
||||
Nur diese Syntax erzeugt Kanten:
|
||||
|
||||
[[rel:depends_on Embeddings 101]]
|
||||
|
||||
Nicht unterstützt:
|
||||
|
||||
rel:depends_on [[Embeddings 101]]
|
||||
|
||||
Dies muss klar dokumentiert werden.
|
||||
|
||||
### 6.2 Prioritäten der Kantenerzeugung
|
||||
Gültige Reihenfolge:
|
||||
|
||||
1. Inline
|
||||
2. Callout
|
||||
3. Typdefaults
|
||||
4. Strukturkanten
|
||||
|
||||
### 6.3 Relaunch der Chunk-zentrierten Funktionslogik
|
||||
Alle Beispiele müssen angepasst werden.
|
||||
|
||||
---
|
||||
|
||||
# 7) chunking_strategy.md
|
||||
## Erforderliche Änderungen
|
||||
|
||||
### 7.1 Neues Verhalten der Chunks
|
||||
- `ord` ist jetzt Sortierkriterium
|
||||
- `window` wird über chunk_config erzeugt
|
||||
- Chunkprofile kommen ausschließlich aus `types.yaml`
|
||||
|
||||
### 7.2 Best Practices erneuern
|
||||
Empfehlung:
|
||||
|
||||
- keine manuelle Chunkprofile mehr im Frontmatter setzen
|
||||
|
||||
---
|
||||
|
||||
# 8) docs_mindnet_retriever.md
|
||||
## Erforderliche Änderungen
|
||||
|
||||
### 8.1 Neue Ausgabeparameter dokumentieren
|
||||
Retriever liefert jetzt:
|
||||
|
||||
semantic_score
|
||||
edge_bonus
|
||||
centrality_bonus
|
||||
type_score
|
||||
total_score
|
||||
paths[]
|
||||
|
||||
### 8.2 Graph-Expansion
|
||||
Dokumentieren:
|
||||
|
||||
- Seed-Auswahl
|
||||
- Depth-Limit
|
||||
- Edge-Filterung
|
||||
- Wirkung von `confidence` auf edge_bonus
|
||||
|
||||
---
|
||||
|
||||
# 9) Handbuch.md
|
||||
## Erforderliche Änderungen
|
||||
|
||||
### 9.1 Neuer Import-Workflow
|
||||
Erläuterung der Parser:
|
||||
|
||||
- Inline-Parser
|
||||
- Callout-Parser
|
||||
- Default-Kanten
|
||||
- Struktur-Kanten
|
||||
|
||||
### 9.2 Best Practices
|
||||
Ergänzen:
|
||||
|
||||
- Typenpflege erfolgt nur in `types.yaml`
|
||||
- Tests immer nach Änderungen ausführen
|
||||
|
||||
---
|
||||
|
||||
# Gesamtübersicht der notwendigen Anpassungen
|
||||
|
||||
| Dokument | Anpassungsbedarf |
|
||||
|---------|------------------|
|
||||
| knowledge_design.md | neue Edge-Arten, Felder, Chunk-Edges |
|
||||
| TYPE_REGISTRY_MANUAL.md | Typauflösung, neue Felder, neue Fallbacks |
|
||||
| wp04_retriever_scoring.md | neue Score-Formel, Pfade, confidence |
|
||||
| implementation_playbook.md | neuer Importprozess, Tests |
|
||||
| technical_architecture.md | aktualisierte Schemas |
|
||||
| functional_architecture.md | Syntax + Prioritäten |
|
||||
| chunking_strategy.md | ord, window, neue Profile |
|
||||
| docs_mindnet_retriever.md | neue API-Parameter |
|
||||
| Handbuch.md | Parser + Best Practices |
|
||||
|
||||
---
|
||||
|
||||
# Empfehlung zur Umsetzung
|
||||
1. Dieses Review-Dokument als Master-Referenz verwenden.
|
||||
2. Jedes Dokument strikt nur an den beschriebenen Stellen anpassen.
|
||||
3. Überprüfen, ob Wann/Wo Typen, Kanten und Parser erklärt werden.
|
||||
4. Nach jeder Anpassung Testlauf:
|
||||
|
||||
python3 -m scripts.payload_dryrun
|
||||
python3 -m scripts.edges_full_check
|
||||
|
||||
5. Danach Playbook + Programmplan updaten.
|
||||
|
||||
@@ -0,0 +1,322 @@
|
||||
# mindnet – Änderungs- und Ergänzungsleitfaden für bestehende Dokumentation
|
||||
# (Stand: Tag Wp04a-Retriever_final)
|
||||
|
||||
Dieser Leitfaden listet **alle notwendigen Änderungen und Ergänzungen** für die
|
||||
bestehenden Dokumente auf. Er umfasst ausschließlich Inhalte, die durch die
|
||||
tatsächliche Implementierung von WP-04a (Retriever, Hybrid Scoring, Graph-Integration,
|
||||
retriever.yaml, Explainability-Vorbereitung, Centrality, Edge-Bonus, Query-API)
|
||||
betroffen sind.
|
||||
|
||||
Er dient dazu, dass ein anderer Chat alle Dokumente vollständig und korrekt
|
||||
überarbeiten kann, ohne Interpretationsspielraum oder fehlende Informationen.
|
||||
|
||||
|
||||
============================================================
|
||||
# 1. Änderungen für: knowledge_design.md
|
||||
============================================================
|
||||
|
||||
## 1.1 Ergänzen: Rolle von retriever_weight pro Note-Type
|
||||
Die Implementierung nutzt `retriever_weight`, das im Importer aus `types.yaml`
|
||||
pro Note-Type übernommen und auf *Chunk-Ebene* vererbt wird.
|
||||
|
||||
**Ergänzen:**
|
||||
- Erklärung, dass `retriever_weight` ein konzeptionelles Typ-Attribut ist,
|
||||
das im hybriden Scoring als Multiplikator wirkt.
|
||||
- Klarstellung, dass `retriever_weight` *nicht* durch Markdown-Frontmatter
|
||||
überschrieben werden muss.
|
||||
- Dokumentation, dass Qdrant-Payload dieses Feld für Notes und Chunks enthält.
|
||||
|
||||
## 1.2 Ergänzen: Edge-Typen, die im Retriever genutzt werden
|
||||
Der Retriever unterstützt aktuell:
|
||||
- `references`
|
||||
- `related_to`
|
||||
- `depends_on`
|
||||
- `similar_to`
|
||||
- `belongs_to`
|
||||
- sowie generierte Strukturkanten `next` / `prev`.
|
||||
|
||||
**Ergänzen:**
|
||||
- Welche Edges für den Graph-Bonus ausgewertet werden.
|
||||
- Wie `confidence` aus mindnet_edges genutzt wird.
|
||||
|
||||
## 1.3 Ergänzen: Definition des lokalen Subgraphen
|
||||
Die reale Retrieval-Logik nutzt:
|
||||
- Seed-Notes (aus semantisch Top-K)
|
||||
- Expansionstiefe `depth`
|
||||
- Edge-Typ-Filter
|
||||
- Bonusberechnung pro Knoten
|
||||
|
||||
Dies ist im ursprünglichen Dokument nicht beschrieben und muss hinzugefügt werden.
|
||||
|
||||
============================================================
|
||||
# 2. Änderungen für: chunking_strategy.md
|
||||
============================================================
|
||||
|
||||
## 2.1 Ergänzen: Chunk-Payload-Felder, die der Retriever nutzt
|
||||
Die aktuelle Strategy beschreibt Chunks korrekt, es fehlt jedoch:
|
||||
|
||||
**Ergänzen:**
|
||||
- `retriever_weight` (vererbt vom Note-Type)
|
||||
- `chunk_profile` (aktuell ungenutzt im Retriever, aber Teil der Payload)
|
||||
- Hinweise, dass der Retriever nur aus *Chunks* semantische Seeds bildet,
|
||||
das Ranking aber auf *Notes* aggregiert wird.
|
||||
|
||||
## 2.2 Ergänzen: Section/Window-Felder und ihre Relevanz
|
||||
Die Felder existieren im Qdrant, werden aber im retriever.yaml nicht verändert.
|
||||
|
||||
**Hinweis ergänzen:**
|
||||
- Der Retriever nutzt aktuell `text` und `window`; kein Einfluss der Abschnittslogik.
|
||||
|
||||
|
||||
============================================================
|
||||
# 3. Änderungen für: TYPE_REGISTRY_MANUAL.md
|
||||
============================================================
|
||||
|
||||
Dieses Dokument muss **aktualisiert** werden, da WP-04a folgende Mechanismen nutzt:
|
||||
|
||||
## 3.1 Ergänzen: Bedeutung von retriever_weight
|
||||
- Definition als quantitatives Typ-Merkmal.
|
||||
- Erklärung, dass es in Qdrant gespeichert wird.
|
||||
- Erklärung des Einflusses auf Semantikscore:
|
||||
- Formel:
|
||||
|
||||
final_semantic = semantic_score * retriever_weight
|
||||
|
||||
- Hinweis, dass in future WP-08 das Gewicht durch Feedback lernbar wird.
|
||||
|
||||
## 3.2 Ergänzen: Neue Edge-Typen und ihre Standard-Konfiguration
|
||||
- `similar_to`
|
||||
- `belongs_to`
|
||||
- Strukturkanten `next` / `prev`
|
||||
|
||||
Diese sind im Importer bereits aktiv, aber nicht ausreichend dokumentiert.
|
||||
|
||||
## 3.3 Ergänzen: Edge-Konfidenzen
|
||||
Die automatische Edge-Ableitung generiert `confidence`.
|
||||
Retriever nutzt diesen Wert linear im Edge-Bonus.
|
||||
|
||||
**Dokumentation muss ergänzt werden.**
|
||||
|
||||
|
||||
============================================================
|
||||
# 4. Änderungen für: docs_mindnet_retriever.md
|
||||
============================================================
|
||||
|
||||
Dieses Dokument ist am stärksten betroffen und muss auf den Stand der tatsächlichen
|
||||
Implementierung aktualisiert werden.
|
||||
|
||||
## 4.1 Aktualisieren: Gesamtscoring-Formel
|
||||
Bisher: Semantik + einfache Typgewichtung.
|
||||
|
||||
Neu implementiert:
|
||||
|
||||
total_score =
|
||||
semantic_weight * semantic_score
|
||||
+ edge_weight * edge_bonus
|
||||
+ centrality_weight * centrality_bonus
|
||||
|
||||
Diese Formel muss vollständig dokumentiert werden.
|
||||
|
||||
## 4.2 Aktualisieren: Einführung retriever.yaml
|
||||
Dokumentation benötigt:
|
||||
|
||||
- Speicherort: `app/config/retriever.yaml`
|
||||
- Default-Werte
|
||||
- Abschnittsweise Beschreibung:
|
||||
- semantic_weight
|
||||
- edge_weight
|
||||
- centrality_weight
|
||||
- Hinweis: Änderungen wirken ohne Codeänderung.
|
||||
|
||||
## 4.3 Ergänzen: Beschreibung Graph-Expansion
|
||||
Der Retriever nutzt:
|
||||
|
||||
- Expand-depth (default 1)
|
||||
- Edge-Typ-Filter
|
||||
- Aggregation pro Node
|
||||
- Bonus-Skalierung
|
||||
|
||||
Diese Logik ist neu und muss detailliert erklärt werden.
|
||||
|
||||
## 4.4 Ergänzen: Beschreibung centrality_bonus
|
||||
Aktuell:
|
||||
- proportionale Verteilung anhand Knotendichte
|
||||
- Normalisierung über lokales Subgraph-Maß
|
||||
|
||||
Dokumentation muss Minimum enthalten:
|
||||
- wie centrality gewichtet wird
|
||||
- welche Grenze existieren
|
||||
|
||||
## 4.5 Ergänzen: Explainability-Schnittstelle (WP-04b Vorbereitung)
|
||||
Die Implementierung erzeugt intern:
|
||||
- edge_bonus pro Node
|
||||
- centrality_bonus pro Node
|
||||
- Liste der Edges, die beteiligt waren
|
||||
|
||||
Diese Werte sind abrufbar und müssen dokumentiert werden.
|
||||
|
||||
## 4.6 Ergänzen: Fehlertoleranzen & Timeout (Health Check)
|
||||
Der Retriever kann:
|
||||
- Fallback auf Vektor-Suche (pure semantic)
|
||||
- Fallback bei fehlendem expand
|
||||
|
||||
Diese Sicherheitsmechanismen müssen dokumentiert werden.
|
||||
|
||||
|
||||
============================================================
|
||||
# 5. Änderungen für: mindnet_technical_architecture.md
|
||||
============================================================
|
||||
|
||||
## 5.1 Ergänzen: Komponentenübersicht Retrieval-Layer
|
||||
Eine neue Abbildung oder Abschnitt muss hinzu:
|
||||
|
||||
- QueryRequest-Parsing
|
||||
- Embedding-Lookup oder Vector-Passthrough
|
||||
- Semantic Seed Search (Qdrant)
|
||||
- Graph-Expansion (mindnet_edges)
|
||||
- Node-/Note-Aggregation
|
||||
- Score-Combination
|
||||
- Response-Ausgabe via FastAPI
|
||||
|
||||
## 5.2 Ergänzen: Datenstruktur mindnet_edges
|
||||
Vor WP-04 nicht vollständig dokumentiert.
|
||||
|
||||
Benötigt:
|
||||
|
||||
- Felder:
|
||||
- edge_id
|
||||
- kind
|
||||
- source_id
|
||||
- target_id
|
||||
- confidence
|
||||
- rule_id
|
||||
- Wie Edges im Retriever ausgewertet werden.
|
||||
|
||||
## 5.3 Ergänzen: retriever.yaml als Konfigurationsquelle
|
||||
Architekturdiagramm muss updated werden:
|
||||
- Retriever nutzt YAML
|
||||
- Keine Hardcodierung im Python
|
||||
|
||||
## 5.4 Ergänzen: Health Check Script
|
||||
Technischer Abschnitt muss dokumentieren:
|
||||
|
||||
- health_check_mindnet.py
|
||||
- Semantik:
|
||||
- semantic query
|
||||
- hybrid query
|
||||
- Fehlercodes
|
||||
|
||||
|
||||
============================================================
|
||||
# 6. Änderungen für: mindnet_functional_architecture.md
|
||||
============================================================
|
||||
|
||||
## 6.1 Ergänzen: Funktionale Rolle des Retrievers
|
||||
Retriever ist jetzt:
|
||||
|
||||
- Teil des Functional Layers
|
||||
- Grundlage für Agenten / Decision Engine
|
||||
- Hauptquelle für Begründungen
|
||||
|
||||
## 6.2 Ergänzen: Einbindung des Explainability-Layers
|
||||
Funktional gehört dies zwischen:
|
||||
- Retriever
|
||||
- Antwortgenerierung (LLM)
|
||||
|
||||
Der Explainability-Layer formuliert Gründe:
|
||||
- Welche Edges?
|
||||
- Welche Scores?
|
||||
- Warum diese Note?
|
||||
|
||||
## 6.3 Ergänzen: Self-Tuning als neuen Funktionsblock (WP-08)
|
||||
Muss aufgenommen werden:
|
||||
|
||||
- Logging von Feedback
|
||||
- Bewertung
|
||||
- Tuning der retriever.yaml
|
||||
- Human-in-the-loop Freigabe
|
||||
|
||||
|
||||
============================================================
|
||||
# 7. Änderungen für: wp04_retriever_scoring.md
|
||||
============================================================
|
||||
|
||||
Dieses Dokument ist **direkt WP-04** und muss am stärksten überarbeitet werden.
|
||||
|
||||
## 7.1 Aktualisieren: Finale Scoring-Formel
|
||||
Die Formel muss exakt so dokumentiert werden, wie inzwischen implementiert.
|
||||
|
||||
## 7.2 Ergänzen: Bonusberechnung
|
||||
Die bisherige Planung wurde implementiert, aber der Text muss aktualisiert werden:
|
||||
|
||||
- Bonusberechnung je Edge
|
||||
- confidence-Einfluss
|
||||
- Node-Aggregation
|
||||
- Note-Aggregation
|
||||
|
||||
## 7.3 Ergänzen: Centrality
|
||||
Dokument ist zu ergänzen um:
|
||||
|
||||
- Definition
|
||||
- Berechnung
|
||||
- Normalisierung
|
||||
- Gewichtung über retriever.yaml
|
||||
|
||||
## 7.4 Ergänzen: Teststrategie
|
||||
Die finale Testbasis muss dokumentiert werden:
|
||||
|
||||
- Unit-Tests
|
||||
- Hybrid-Tests
|
||||
- API-Tests
|
||||
- Smoke-Test
|
||||
- Diagnose-Skripte
|
||||
|
||||
## 7.5 Ergänzen: Einschränkungen & Offene Punkte
|
||||
Hinweis für WP-04b:
|
||||
- Explainability fehlt noch
|
||||
- Edge-Type Weights noch nicht implementiert
|
||||
- Self-Tuning in WP-08
|
||||
|
||||
|
||||
============================================================
|
||||
# 8. Änderungen für: mindnet_v2_implementation_playbook.md
|
||||
============================================================
|
||||
|
||||
## 8.1 Ergänzen: Arbeitsablauf mit Retriever
|
||||
Einrichtung, Nutzung und Test sollte ergänzt werden:
|
||||
|
||||
- Starten von FastAPI
|
||||
- Testen des Retrievers via health_check
|
||||
- Import-Workflow (neue Vaults → Qdrant → Retriever)
|
||||
|
||||
## 8.2 Ergänzen: Konfigurationsverwaltung
|
||||
Einbindung in Gitea-Workflows:
|
||||
- retriever.yaml versionieren
|
||||
- Tags/Branches für Funktionsstände setzen
|
||||
|
||||
## 8.3 Ergänzen: Self-Tuning-Prozess (nur Rahmen)
|
||||
Der Playbook-Abschnitt soll die späteren Mechanismen vorbeschreiben.
|
||||
|
||||
|
||||
============================================================
|
||||
# 9. Änderungen für: Handbuch.md
|
||||
============================================================
|
||||
|
||||
## 9.1 Ergänzen: Einfache Erklärung des Retrievers
|
||||
Das Handbuch benötigt eine verständliche, nicht-technische Beschreibung:
|
||||
|
||||
- Was macht der Retriever?
|
||||
- Warum kommt ein Ergebnis zustande?
|
||||
- Welche Rolle spielen Edges?
|
||||
- Wie kann der Nutzer die Gewichtung beeinflussen (YAML)?
|
||||
|
||||
## 9.2 Ergänzen: Troubleshooting
|
||||
Aufnehmen:
|
||||
- „Ergebnisse wirken unpassend“ → Prüfe retriever.yaml
|
||||
- „Edge-Bonus fehlt“ → Prüfe mindnet_edges
|
||||
- health_check-Skript nutzen
|
||||
|
||||
## 9.3 Ergänzen: API-Nutzung für Endanwender
|
||||
Einfacher Abschnitt:
|
||||
- Wie man `/query` testet
|
||||
- Beispiele für semantic & hybrid queries
|
||||
@@ -33,3 +33,7 @@ python3 scripts/import_markdown.py --vault /path/to/Obsidian
|
||||
- `POST /qdrant/query` → semantic search over chunks with optional filters
|
||||
|
||||
See `scripts/quick_test.sh` for a runnable example.
|
||||
|
||||
>Anmerkung:
|
||||
Diese Datei ist veraltet
|
||||
|
||||
+111
-162
@@ -1,180 +1,129 @@
|
||||
# chunk_payload.py
|
||||
#!/usr/bin/env python3
|
||||
# -*- coding: utf-8 -*-
|
||||
"""
|
||||
app/core/chunk_payload.py (Mindnet V2 — types.yaml authoritative)
|
||||
- neighbors_prev / neighbors_next sind Listen ([], [id]).
|
||||
- retriever_weight / chunk_profile kommen aus types.yaml (Frontmatter wird ignoriert).
|
||||
- Fallbacks: defaults.* in types.yaml; sonst 1.0 / "default".
|
||||
"""
|
||||
from __future__ import annotations
|
||||
from typing import Any, Dict, List, Optional, Tuple
|
||||
import os, json, pathlib, re, yaml, hashlib
|
||||
from typing import Any, Dict, List, Optional
|
||||
import os, yaml
|
||||
|
||||
FRONTMATTER_RE = re.compile(r"^---\s*\n.*?\n---\s*\n?", re.DOTALL)
|
||||
def _env(n: str, d: Optional[str]=None) -> str:
|
||||
v = os.getenv(n)
|
||||
return v if v is not None else (d or "")
|
||||
|
||||
def _as_dict(note: Any) -> Dict[str, Any]:
|
||||
if isinstance(note, dict):
|
||||
return note
|
||||
d: Dict[str, Any] = {}
|
||||
for attr in ("id","note_id","title","path","frontmatter","meta","metadata","body","text","content","raw","markdown","type","chunks"):
|
||||
if hasattr(note, attr):
|
||||
d[attr] = getattr(note, attr)
|
||||
fm = d.get("frontmatter") or d.get("meta") or d.get("metadata") or {}
|
||||
d["frontmatter"] = fm if isinstance(fm, dict) else {}
|
||||
return d
|
||||
def _load_types() -> dict:
|
||||
p = _env("MINDNET_TYPES_FILE", "./config/types.yaml")
|
||||
try:
|
||||
with open(p, "r", encoding="utf-8") as f:
|
||||
return yaml.safe_load(f) or {}
|
||||
except Exception:
|
||||
return {}
|
||||
|
||||
def _pick_args(*args, **kwargs) -> Tuple[Optional[str], Optional[Dict[str,Any]]]:
|
||||
path = kwargs.get("path")
|
||||
types_cfg = kwargs.get("types_config")
|
||||
for a in args:
|
||||
if path is None and isinstance(a, (str, pathlib.Path)):
|
||||
path = str(a)
|
||||
if types_cfg is None and isinstance(a, dict):
|
||||
types_cfg = a
|
||||
return path, types_cfg
|
||||
def _get_types_map(reg: dict) -> dict:
|
||||
if isinstance(reg, dict) and isinstance(reg.get("types"), dict):
|
||||
return reg["types"]
|
||||
return reg if isinstance(reg, dict) else {}
|
||||
|
||||
def _load_types_config(explicit: Optional[Dict[str, Any]] = None) -> Dict[str, Any]:
|
||||
if isinstance(explicit, dict):
|
||||
return explicit
|
||||
for rel in ("config/config.yaml", "config/types.yaml"):
|
||||
p = pathlib.Path(rel)
|
||||
if p.exists():
|
||||
with p.open("r", encoding="utf-8") as f:
|
||||
data = yaml.safe_load(f) or {}
|
||||
if isinstance(data, dict) and "types" in data and isinstance(data["types"], dict):
|
||||
return data["types"]
|
||||
return data if isinstance(data, dict) else {}
|
||||
def _get_defaults(reg: dict) -> dict:
|
||||
if isinstance(reg, dict) and isinstance(reg.get("defaults"), dict):
|
||||
return reg["defaults"]
|
||||
if isinstance(reg, dict) and isinstance(reg.get("global"), dict):
|
||||
return reg["global"]
|
||||
return {}
|
||||
|
||||
def _coalesce(*vals):
|
||||
for v in vals:
|
||||
def _as_float(x: Any):
|
||||
try:
|
||||
return float(x)
|
||||
except Exception:
|
||||
return None
|
||||
|
||||
def _resolve_chunk_profile(note_type: str, reg: dict) -> str:
|
||||
types = _get_types_map(reg)
|
||||
if isinstance(types, dict):
|
||||
t = types.get(note_type, {})
|
||||
if isinstance(t, dict) and isinstance(t.get("chunk_profile"), str):
|
||||
return t["chunk_profile"]
|
||||
defs = _get_defaults(reg)
|
||||
if isinstance(defs, dict) and isinstance(defs.get("chunk_profile"), str):
|
||||
return defs["chunk_profile"]
|
||||
return "default"
|
||||
|
||||
def _resolve_retriever_weight(note_type: str, reg: dict) -> float:
|
||||
types = _get_types_map(reg)
|
||||
if isinstance(types, dict):
|
||||
t = types.get(note_type, {})
|
||||
if isinstance(t, dict) and (t.get("retriever_weight") is not None):
|
||||
v = _as_float(t.get("retriever_weight"))
|
||||
if v is not None:
|
||||
return float(v)
|
||||
defs = _get_defaults(reg)
|
||||
if isinstance(defs, dict) and (defs.get("retriever_weight") is not None):
|
||||
v = _as_float(defs.get("retriever_weight"))
|
||||
if v is not None:
|
||||
return v
|
||||
return None
|
||||
return float(v)
|
||||
return 1.0
|
||||
|
||||
def _text_from_note(n: Dict[str, Any], path_hint: Optional[str]) -> str:
|
||||
# häufige Felder
|
||||
cand = [
|
||||
n.get("body"),
|
||||
n.get("text"),
|
||||
n.get("markdown"),
|
||||
n.get("raw"),
|
||||
]
|
||||
content = n.get("content")
|
||||
if isinstance(content, str):
|
||||
cand.append(content)
|
||||
elif isinstance(content, dict):
|
||||
for k in ("text","body","raw","markdown","content"):
|
||||
v = content.get(k)
|
||||
if isinstance(v, str):
|
||||
cand.append(v)
|
||||
for t in cand:
|
||||
if isinstance(t, str) and t.strip():
|
||||
return t
|
||||
|
||||
# Fallback: Datei lesen und Frontmatter entfernen
|
||||
p = n.get("path") or path_hint
|
||||
if p:
|
||||
try:
|
||||
pth = pathlib.Path(p)
|
||||
if pth.exists():
|
||||
txt = pth.read_text(encoding="utf-8", errors="ignore")
|
||||
if txt:
|
||||
return FRONTMATTER_RE.sub("", txt).strip()
|
||||
except Exception:
|
||||
pass
|
||||
return ""
|
||||
|
||||
def _iter_chunks(n: Dict[str, Any], profile: str, fulltext: str) -> List[Dict[str, Any]]:
|
||||
# 1) vorhandene Chunks nehmen, wenn sinnvoll
|
||||
existing = n.get("chunks")
|
||||
out: List[Dict[str, Any]] = []
|
||||
if isinstance(existing, list) and existing:
|
||||
for i, c in enumerate(existing):
|
||||
text = ""
|
||||
if isinstance(c, dict):
|
||||
text = c.get("text") or c.get("body") or c.get("raw") or ""
|
||||
elif isinstance(c, str):
|
||||
text = c
|
||||
if text and text.strip():
|
||||
out.append({"index": i, "text": text})
|
||||
if out:
|
||||
return out
|
||||
|
||||
# 2) Fallback: profilabhängige Bündelung
|
||||
if not isinstance(profile, str):
|
||||
profile = "medium"
|
||||
size = {"short": 600, "medium": 1200, "long": 2400}.get(profile, 1200)
|
||||
if not fulltext:
|
||||
def _as_list(x):
|
||||
if x is None:
|
||||
return []
|
||||
paras = re.split(r"\n{2,}", fulltext)
|
||||
buf = ""
|
||||
chunks: List[str] = []
|
||||
for p in paras:
|
||||
p = (p or "").strip()
|
||||
if not p:
|
||||
continue
|
||||
if len(buf) + (2 if buf else 0) + len(p) <= size:
|
||||
buf = (buf + "\n\n" + p).strip() if buf else p
|
||||
else:
|
||||
if buf: chunks.append(buf)
|
||||
if len(p) <= size:
|
||||
buf = p
|
||||
else:
|
||||
for i in range(0, len(p), size):
|
||||
chunks.append(p[i:i+size])
|
||||
buf = ""
|
||||
if buf: chunks.append(buf)
|
||||
return [{"index": i, "text": c} for i, c in enumerate(chunks)]
|
||||
if isinstance(x, list):
|
||||
return x
|
||||
return [x]
|
||||
|
||||
def make_chunk_payloads(note: Any, *args, **kwargs) -> List[Dict[str, Any]]:
|
||||
n = _as_dict(note)
|
||||
path_arg, types_cfg_explicit = _pick_args(*args, **kwargs)
|
||||
types_cfg = _load_types_config(types_cfg_explicit)
|
||||
def make_chunk_payloads(note: Dict[str, Any],
|
||||
note_path: str,
|
||||
chunks_from_chunker: List[Any],
|
||||
*,
|
||||
note_text: str = "",
|
||||
types_cfg: Optional[dict] = None,
|
||||
file_path: Optional[str] = None) -> List[Dict[str, Any]]:
|
||||
fm = (note or {}).get("frontmatter", {}) or {}
|
||||
note_type = fm.get("type") or note.get("type") or "concept"
|
||||
reg = types_cfg if isinstance(types_cfg, dict) else _load_types()
|
||||
|
||||
fm = n.get("frontmatter") or {}
|
||||
note_type = str(fm.get("type") or n.get("type") or "note")
|
||||
cfg_for_type = types_cfg.get(note_type, {}) if isinstance(types_cfg, dict) else {}
|
||||
# types.yaml authoritative
|
||||
cp = _resolve_chunk_profile(note_type, reg)
|
||||
rw = _resolve_retriever_weight(note_type, reg)
|
||||
|
||||
try:
|
||||
default_rw = float(os.environ.get("MINDNET_DEFAULT_RETRIEVER_WEIGHT", 1.0))
|
||||
except Exception:
|
||||
default_rw = 1.0
|
||||
tags = fm.get("tags") or []
|
||||
if isinstance(tags, str):
|
||||
tags = [tags]
|
||||
|
||||
retriever_weight = _coalesce(fm.get("retriever_weight"), cfg_for_type.get("retriever_weight"), default_rw)
|
||||
try:
|
||||
retriever_weight = float(retriever_weight)
|
||||
except Exception:
|
||||
retriever_weight = default_rw
|
||||
out: List[Dict[str, Any]] = []
|
||||
for idx, ch in enumerate(chunks_from_chunker):
|
||||
# Attribute oder Keys (Chunk-Objekt oder Dict)
|
||||
cid = getattr(ch, "id", None) or (ch.get("id") if isinstance(ch, dict) else None)
|
||||
nid = getattr(ch, "note_id", None) or (ch.get("note_id") if isinstance(ch, dict) else fm.get("id"))
|
||||
index = getattr(ch, "index", None) or (ch.get("index") if isinstance(ch, dict) else idx)
|
||||
text = getattr(ch, "text", None) or (ch.get("text") if isinstance(ch, dict) else "")
|
||||
window = getattr(ch, "window", None) or (ch.get("window") if isinstance(ch, dict) else text)
|
||||
prev_id = getattr(ch, "neighbors_prev", None) or (ch.get("neighbors_prev") if isinstance(ch, dict) else None)
|
||||
next_id = getattr(ch, "neighbors_next", None) or (ch.get("neighbors_next") if isinstance(ch, dict) else None)
|
||||
|
||||
chunk_profile = _coalesce(fm.get("chunk_profile"), cfg_for_type.get("chunk_profile"), os.environ.get("MINDNET_DEFAULT_CHUNK_PROFILE","medium"))
|
||||
chunk_profile = chunk_profile if isinstance(chunk_profile, str) else "medium"
|
||||
|
||||
note_id = n.get("note_id") or n.get("id") or fm.get("id")
|
||||
title = n.get("title") or fm.get("title") or ""
|
||||
path = n.get("path") or path_arg
|
||||
if isinstance(path, pathlib.Path):
|
||||
path = str(path)
|
||||
path = path or "" # immer vorhanden
|
||||
|
||||
fulltext = _text_from_note(n, path)
|
||||
chunks = _iter_chunks(n, chunk_profile, fulltext)
|
||||
|
||||
payloads: List[Dict[str, Any]] = []
|
||||
for c in chunks:
|
||||
idx = c.get("index", len(payloads))
|
||||
text = c.get("text") if isinstance(c, dict) else (str(c) if c is not None else "")
|
||||
text = text if isinstance(text, str) else str(text or "")
|
||||
|
||||
key = f"{note_id}|{idx}"
|
||||
h = hashlib.sha1(key.encode("utf-8")).hexdigest()[:12] if note_id else hashlib.sha1(f"{path}|{idx}".encode("utf-8")).hexdigest()[:12]
|
||||
chunk_id = f"{note_id}-{idx:03d}-{h}" if note_id else f"{h}"
|
||||
|
||||
payload = {
|
||||
"note_id": note_id,
|
||||
"chunk_id": chunk_id,
|
||||
"index": idx,
|
||||
"title": title,
|
||||
pl: Dict[str, Any] = {
|
||||
"note_id": nid,
|
||||
"chunk_id": cid,
|
||||
"index": int(index),
|
||||
"ord": int(index) + 1,
|
||||
"type": note_type,
|
||||
"path": path, # <- garantiert vorhanden
|
||||
"text": text, # <- nie leer, sonst werden keine Chunks erzeugt
|
||||
"retriever_weight": retriever_weight,
|
||||
"chunk_profile": chunk_profile,
|
||||
"tags": tags,
|
||||
"text": text,
|
||||
"window": window,
|
||||
"neighbors_prev": _as_list(prev_id),
|
||||
"neighbors_next": _as_list(next_id),
|
||||
"section": getattr(ch, "section", None) or (ch.get("section") if isinstance(ch, dict) else ""),
|
||||
"path": note_path,
|
||||
"source_path": file_path or note_path,
|
||||
"retriever_weight": float(rw),
|
||||
"chunk_profile": cp,
|
||||
}
|
||||
json.loads(json.dumps(payload, ensure_ascii=False))
|
||||
payloads.append(payload)
|
||||
# Aufräumen von Alt-Feldern
|
||||
for alias in ("chunk_num", "Chunk_Number"):
|
||||
pl.pop(alias, None)
|
||||
|
||||
return payloads
|
||||
out.append(pl)
|
||||
|
||||
return out
|
||||
+354
-50
@@ -2,43 +2,45 @@
|
||||
# -*- coding: utf-8 -*-
|
||||
"""
|
||||
Modul: app/core/derive_edges.py
|
||||
Version: 1.4.0
|
||||
Datum: 2025-10-01
|
||||
|
||||
Zweck
|
||||
-----
|
||||
Robuste Kantenbildung für mindnet (Notes/Chunks):
|
||||
- belongs_to (chunk -> note)
|
||||
- next / prev (chunk-Kette)
|
||||
- references (chunk-scope) aus Chunk.window/text
|
||||
- optional references/backlink (note-scope)
|
||||
|
||||
Wichtig: Wikilinks werden mit der Parser-Funktion `extract_wikilinks` extrahiert,
|
||||
damit Varianten wie [[id#anchor]] oder [[id|label]] korrekt auf 'id' reduziert werden.
|
||||
|
||||
Erwartete Chunk-Payload-Felder:
|
||||
{
|
||||
"note_id": "...",
|
||||
"chunk_id": "...", # Alias "id" ist zulässig
|
||||
"id": "...",
|
||||
"chunk_index": int,
|
||||
"seq": int,
|
||||
"window": str,
|
||||
"text": str,
|
||||
"path": "rel/path.md",
|
||||
...
|
||||
}
|
||||
Zweck:
|
||||
- Bewahrt bestehende Edgelogik (belongs_to, prev/next, references, backlink)
|
||||
- Ergänzt typenbasierte Default-Kanten (edge_defaults aus config/types.yaml)
|
||||
- Unterstützt "typed inline relations":
|
||||
* [[rel:KIND | Target]]
|
||||
* [[rel:KIND Target]]
|
||||
* rel: KIND [[Target]]
|
||||
- Unterstützt Obsidian-Callouts:
|
||||
* > [!edge] KIND: [[Target]] [[Target2]] ...
|
||||
Kompatibilität:
|
||||
- build_edges_for_note(...) Signatur unverändert
|
||||
- rule_id Werte:
|
||||
* structure:belongs_to
|
||||
* structure:order
|
||||
* explicit:wikilink
|
||||
* inline:rel
|
||||
* callout:edge
|
||||
* edge_defaults:<type>:<relation>
|
||||
* derived:backlink
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from typing import Dict, List, Optional, Iterable
|
||||
import os
|
||||
import re
|
||||
from typing import Iterable, List, Optional, Tuple, Set, Dict
|
||||
|
||||
# WICHTIG: benutze die Parser-Extraktion für saubere Wikilinks
|
||||
from app.core.parser import extract_wikilinks
|
||||
try:
|
||||
import yaml # optional, nur für types.yaml
|
||||
except Exception: # pragma: no cover
|
||||
yaml = None
|
||||
|
||||
# --------------------------------------------------------------------------- #
|
||||
# Utilities
|
||||
# --------------------------------------------------------------------------- #
|
||||
|
||||
def _get(d: dict, *keys, default=None):
|
||||
for k in keys:
|
||||
if k in d and d[k] is not None:
|
||||
if isinstance(d, dict) and k in d and d[k] is not None:
|
||||
return d[k]
|
||||
return default
|
||||
|
||||
@@ -52,8 +54,8 @@ def _chunk_text_for_refs(chunk: dict) -> str:
|
||||
or ""
|
||||
)
|
||||
|
||||
def _dedupe(seq: Iterable[str]) -> List[str]:
|
||||
seen = set()
|
||||
def _dedupe_seq(seq: Iterable[str]) -> List[str]:
|
||||
seen: Set[str] = set()
|
||||
out: List[str] = []
|
||||
for s in seq:
|
||||
if s not in seen:
|
||||
@@ -64,15 +66,177 @@ def _dedupe(seq: Iterable[str]) -> List[str]:
|
||||
def _edge(kind: str, scope: str, source_id: str, target_id: str, note_id: str, extra: Optional[dict] = None) -> dict:
|
||||
pl = {
|
||||
"kind": kind,
|
||||
"scope": scope, # "chunk" | "note"
|
||||
"relation": kind, # Alias (v2)
|
||||
"scope": scope, # "chunk" | "note"
|
||||
"source_id": source_id,
|
||||
"target_id": target_id,
|
||||
"note_id": note_id, # Träger/Quelle der Kante (aktuelle Note)
|
||||
"note_id": note_id, # Träger-Note der Kante
|
||||
}
|
||||
if extra:
|
||||
pl.update(extra)
|
||||
return pl
|
||||
|
||||
def _mk_edge_id(kind: str, s: str, t: str, scope: str, rule_id: Optional[str] = None) -> str:
|
||||
base = f"{kind}:{s}->{t}#{scope}"
|
||||
if rule_id:
|
||||
base += f"|{rule_id}"
|
||||
try:
|
||||
import hashlib
|
||||
return hashlib.blake2s(base.encode("utf-8"), digest_size=12).hexdigest()
|
||||
except Exception: # pragma: no cover
|
||||
return base
|
||||
|
||||
# --------------------------------------------------------------------------- #
|
||||
# Typen-Registry (types.yaml)
|
||||
# --------------------------------------------------------------------------- #
|
||||
|
||||
def _env(n: str, default: Optional[str] = None) -> str:
|
||||
v = os.getenv(n)
|
||||
return v if v is not None else (default or "")
|
||||
|
||||
def _load_types_registry() -> dict:
|
||||
"""Lädt die YAML-Registry aus MINDNET_TYPES_FILE oder ./config/types.yaml"""
|
||||
p = _env("MINDNET_TYPES_FILE", "./config/types.yaml")
|
||||
if not os.path.isfile(p) or yaml is None:
|
||||
return {}
|
||||
try:
|
||||
with open(p, "r", encoding="utf-8") as f:
|
||||
data = yaml.safe_load(f) or {}
|
||||
return data
|
||||
except Exception:
|
||||
return {}
|
||||
|
||||
def _get_types_map(reg: dict) -> dict:
|
||||
if isinstance(reg, dict) and isinstance(reg.get("types"), dict):
|
||||
return reg["types"]
|
||||
return reg if isinstance(reg, dict) else {}
|
||||
|
||||
def _edge_defaults_for(note_type: Optional[str], reg: dict) -> List[str]:
|
||||
"""
|
||||
Liefert die edge_defaults-Liste für den gegebenen Notiztyp.
|
||||
Fallback-Reihenfolge:
|
||||
1) reg['types'][note_type]['edge_defaults']
|
||||
2) reg['defaults']['edge_defaults'] (oder 'default'/'global')
|
||||
3) []
|
||||
"""
|
||||
types_map = _get_types_map(reg)
|
||||
if note_type and isinstance(types_map, dict):
|
||||
t = types_map.get(note_type)
|
||||
if isinstance(t, dict) and isinstance(t.get("edge_defaults"), list):
|
||||
return [str(x) for x in t["edge_defaults"] if isinstance(x, str)]
|
||||
for key in ("defaults", "default", "global"):
|
||||
v = reg.get(key)
|
||||
if isinstance(v, dict) and isinstance(v.get("edge_defaults"), list):
|
||||
return [str(x) for x in v["edge_defaults"] if isinstance(x, str)]
|
||||
return []
|
||||
|
||||
# --------------------------------------------------------------------------- #
|
||||
# Parser für Links / Relationen
|
||||
# --------------------------------------------------------------------------- #
|
||||
|
||||
# Normale Wikilinks (Fallback)
|
||||
_WIKILINK_RE = re.compile(r"\[\[(?:[^\|\]]+\|)?([a-zA-Z0-9_\-#:. ]+)\]\]")
|
||||
|
||||
# Getypte Inline-Relationen:
|
||||
# [[rel:KIND | Target]]
|
||||
# [[rel:KIND Target]]
|
||||
_REL_PIPE = re.compile(r"\[\[\s*rel:(?P<kind>[a-z_]+)\s*\|\s*(?P<target>[^\]]+?)\s*\]\]", re.IGNORECASE)
|
||||
_REL_SPACE = re.compile(r"\[\[\s*rel:(?P<kind>[a-z_]+)\s+(?P<target>[^\]]+?)\s*\]\]", re.IGNORECASE)
|
||||
# rel: KIND [[Target]] (reines Textmuster)
|
||||
_REL_TEXT = re.compile(r"rel\s*:\s*(?P<kind>[a-z_]+)\s*\[\[\s*(?P<target>[^\]]+?)\s*\]\]", re.IGNORECASE)
|
||||
|
||||
def _extract_typed_relations(text: str) -> Tuple[List[Tuple[str,str]], str]:
|
||||
"""
|
||||
Gibt Liste (kind, target) zurück und den Text mit entfernten getypten Relation-Links,
|
||||
damit die generische Wikilink-Erkennung sie nicht doppelt zählt.
|
||||
Unterstützt drei Varianten:
|
||||
- [[rel:KIND | Target]]
|
||||
- [[rel:KIND Target]]
|
||||
- rel: KIND [[Target]]
|
||||
"""
|
||||
pairs: List[Tuple[str,str]] = []
|
||||
def _collect(m):
|
||||
k = (m.group("kind") or "").strip().lower()
|
||||
t = (m.group("target") or "").strip()
|
||||
if k and t:
|
||||
pairs.append((k, t))
|
||||
return "" # Link entfernen
|
||||
|
||||
text = _REL_PIPE.sub(_collect, text)
|
||||
text = _REL_SPACE.sub(_collect, text)
|
||||
text = _REL_TEXT.sub(_collect, text)
|
||||
return pairs, text
|
||||
|
||||
# Obsidian Callout Parser
|
||||
_CALLOUT_START = re.compile(r"^\s*>\s*\[!edge\]\s*(.*)$", re.IGNORECASE)
|
||||
_REL_LINE = re.compile(r"^(?P<kind>[a-z_]+)\s*:\s*(?P<targets>.+?)\s*$", re.IGNORECASE)
|
||||
_WIKILINKS_IN_LINE = re.compile(r"\[\[([^\]]+)\]\]")
|
||||
|
||||
def _extract_callout_relations(text: str) -> Tuple[List[Tuple[str,str]], str]:
|
||||
"""
|
||||
Findet [!edge]-Callouts und extrahiert (kind, target). Entfernt den gesamten
|
||||
Callout-Block aus dem Text (damit Wikilinks daraus nicht zusätzlich als
|
||||
"references" gezählt werden).
|
||||
"""
|
||||
if not text:
|
||||
return [], text
|
||||
|
||||
lines = text.splitlines()
|
||||
out_pairs: List[Tuple[str,str]] = []
|
||||
keep_lines: List[str] = []
|
||||
i = 0
|
||||
|
||||
while i < len(lines):
|
||||
m = _CALLOUT_START.match(lines[i])
|
||||
if not m:
|
||||
keep_lines.append(lines[i])
|
||||
i += 1
|
||||
continue
|
||||
|
||||
block_lines: List[str] = []
|
||||
first_rest = m.group(1) or ""
|
||||
if first_rest.strip():
|
||||
block_lines.append(first_rest)
|
||||
|
||||
i += 1
|
||||
while i < len(lines) and lines[i].lstrip().startswith('>'):
|
||||
block_lines.append(lines[i].lstrip()[1:].lstrip())
|
||||
i += 1
|
||||
|
||||
for bl in block_lines:
|
||||
mrel = _REL_LINE.match(bl)
|
||||
if not mrel:
|
||||
continue
|
||||
kind = (mrel.group("kind") or "").strip().lower()
|
||||
targets = mrel.group("targets") or ""
|
||||
found = _WIKILINKS_IN_LINE.findall(targets)
|
||||
if found:
|
||||
for t in found:
|
||||
t = t.strip()
|
||||
if t:
|
||||
out_pairs.append((kind, t))
|
||||
else:
|
||||
for raw in re.split(r"[,;]", targets):
|
||||
t = raw.strip()
|
||||
if t:
|
||||
out_pairs.append((kind, t))
|
||||
|
||||
# Callout wird NICHT in keep_lines übernommen
|
||||
continue
|
||||
|
||||
remainder = "\n".join(keep_lines)
|
||||
return out_pairs, remainder
|
||||
|
||||
def _extract_wikilinks(text: str) -> List[str]:
|
||||
ids: List[str] = []
|
||||
for m in _WIKILINK_RE.finditer(text or ""):
|
||||
ids.append(m.group(1).strip())
|
||||
return ids
|
||||
|
||||
# --------------------------------------------------------------------------- #
|
||||
# Hauptfunktion
|
||||
# --------------------------------------------------------------------------- #
|
||||
|
||||
def build_edges_for_note(
|
||||
note_id: str,
|
||||
chunks: List[dict],
|
||||
@@ -84,48 +248,188 @@ def build_edges_for_note(
|
||||
|
||||
- belongs_to: für jeden Chunk (chunk -> note)
|
||||
- next / prev: zwischen aufeinanderfolgenden Chunks
|
||||
- references: pro Chunk aus window/text
|
||||
- references: pro Chunk aus window/text (via Wikilinks)
|
||||
- typed inline relations: [[rel:KIND | Target]] / [[rel:KIND Target]] / rel: KIND [[Target]]
|
||||
- Obsidian Callouts: > [!edge] KIND: [[Target]] [[Target2]]
|
||||
- optional note-scope references/backlinks: dedupliziert über alle Chunk-Funde + note_level_references
|
||||
- typenbasierte Default-Kanten (edge_defaults) je gefundener Referenz
|
||||
"""
|
||||
edges: List[dict] = []
|
||||
|
||||
# belongs_to
|
||||
# Note-Typ (aus erstem Chunk erwartet)
|
||||
note_type = None
|
||||
if chunks:
|
||||
note_type = _get(chunks[0], "type")
|
||||
|
||||
# 1) belongs_to
|
||||
for ch in chunks:
|
||||
cid = _get(ch, "chunk_id", "id")
|
||||
if not cid:
|
||||
continue
|
||||
edges.append(_edge("belongs_to", "chunk", cid, note_id, note_id, {"chunk_id": cid}))
|
||||
edges.append(_edge("belongs_to", "chunk", cid, note_id, note_id, {
|
||||
"chunk_id": cid,
|
||||
"edge_id": _mk_edge_id("belongs_to", cid, note_id, "chunk", "structure:belongs_to"),
|
||||
"provenance": "rule",
|
||||
"rule_id": "structure:belongs_to",
|
||||
"confidence": 1.0,
|
||||
}))
|
||||
|
||||
# next/prev
|
||||
# 2) next / prev
|
||||
for i in range(len(chunks) - 1):
|
||||
a, b = chunks[i], chunks[i + 1]
|
||||
a_id = _get(a, "chunk_id", "id")
|
||||
b_id = _get(b, "chunk_id", "id")
|
||||
if not a_id or not b_id:
|
||||
continue
|
||||
edges.append(_edge("next", "chunk", a_id, b_id, note_id, {"chunk_id": a_id}))
|
||||
edges.append(_edge("prev", "chunk", b_id, a_id, note_id, {"chunk_id": b_id}))
|
||||
edges.append(_edge("next", "chunk", a_id, b_id, note_id, {
|
||||
"chunk_id": a_id,
|
||||
"edge_id": _mk_edge_id("next", a_id, b_id, "chunk", "structure:order"),
|
||||
"provenance": "rule",
|
||||
"rule_id": "structure:order",
|
||||
"confidence": 0.95,
|
||||
}))
|
||||
edges.append(_edge("prev", "chunk", b_id, a_id, note_id, {
|
||||
"chunk_id": b_id,
|
||||
"edge_id": _mk_edge_id("prev", b_id, a_id, "chunk", "structure:order"),
|
||||
"provenance": "rule",
|
||||
"rule_id": "structure:order",
|
||||
"confidence": 0.95,
|
||||
}))
|
||||
|
||||
# references (chunk-scope) – Links aus window bevorzugen (Overlap-fest)
|
||||
# 3) references + typed inline + callouts + defaults (chunk-scope)
|
||||
reg = _load_types_registry()
|
||||
defaults = _edge_defaults_for(note_type, reg)
|
||||
refs_all: List[str] = []
|
||||
|
||||
for ch in chunks:
|
||||
cid = _get(ch, "chunk_id", "id")
|
||||
if not cid:
|
||||
continue
|
||||
txt = _chunk_text_for_refs(ch)
|
||||
refs = extract_wikilinks(txt) # <— Parser-Logik, kompatibel zu deinem System
|
||||
raw = _chunk_text_for_refs(ch)
|
||||
|
||||
# 3a) typed inline relations
|
||||
typed, remainder = _extract_typed_relations(raw)
|
||||
for kind, target in typed:
|
||||
kind = kind.strip().lower()
|
||||
if not kind or not target:
|
||||
continue
|
||||
edges.append(_edge(kind, "chunk", cid, target, note_id, {
|
||||
"chunk_id": cid,
|
||||
"edge_id": _mk_edge_id(kind, cid, target, "chunk", "inline:rel"),
|
||||
"provenance": "explicit",
|
||||
"rule_id": "inline:rel",
|
||||
"confidence": 0.95,
|
||||
}))
|
||||
if kind in {"related_to", "similar_to"}:
|
||||
edges.append(_edge(kind, "chunk", target, cid, note_id, {
|
||||
"chunk_id": cid,
|
||||
"edge_id": _mk_edge_id(kind, target, cid, "chunk", "inline:rel"),
|
||||
"provenance": "explicit",
|
||||
"rule_id": "inline:rel",
|
||||
"confidence": 0.95,
|
||||
}))
|
||||
|
||||
# 3b) callouts
|
||||
call_pairs, remainder2 = _extract_callout_relations(remainder)
|
||||
for kind, target in call_pairs:
|
||||
k = (kind or "").strip().lower()
|
||||
if not k or not target:
|
||||
continue
|
||||
edges.append(_edge(k, "chunk", cid, target, note_id, {
|
||||
"chunk_id": cid,
|
||||
"edge_id": _mk_edge_id(k, cid, target, "chunk", "callout:edge"),
|
||||
"provenance": "explicit",
|
||||
"rule_id": "callout:edge",
|
||||
"confidence": 0.95,
|
||||
}))
|
||||
if k in {"related_to", "similar_to"}:
|
||||
edges.append(_edge(k, "chunk", target, cid, note_id, {
|
||||
"chunk_id": cid,
|
||||
"edge_id": _mk_edge_id(k, target, cid, "chunk", "callout:edge"),
|
||||
"provenance": "explicit",
|
||||
"rule_id": "callout:edge",
|
||||
"confidence": 0.95,
|
||||
}))
|
||||
|
||||
# 3c) generische Wikilinks → references (+ defaults je Ref)
|
||||
refs = _extract_wikilinks(remainder2)
|
||||
for r in refs:
|
||||
edges.append(_edge("references", "chunk", cid, r, note_id, {"chunk_id": cid, "ref_text": r}))
|
||||
edges.append(_edge("references", "chunk", cid, r, note_id, {
|
||||
"chunk_id": cid,
|
||||
"ref_text": r,
|
||||
"edge_id": _mk_edge_id("references", cid, r, "chunk", "explicit:wikilink"),
|
||||
"provenance": "explicit",
|
||||
"rule_id": "explicit:wikilink",
|
||||
"confidence": 1.0,
|
||||
}))
|
||||
for rel in defaults:
|
||||
if rel == "references":
|
||||
continue
|
||||
edges.append(_edge(rel, "chunk", cid, r, note_id, {
|
||||
"chunk_id": cid,
|
||||
"edge_id": _mk_edge_id(rel, cid, r, "chunk", f"edge_defaults:{note_type}:{rel}"),
|
||||
"provenance": "rule",
|
||||
"rule_id": f"edge_defaults:{note_type}:{rel}",
|
||||
"confidence": 0.7,
|
||||
}))
|
||||
if rel in {"related_to", "similar_to"}:
|
||||
edges.append(_edge(rel, "chunk", r, cid, note_id, {
|
||||
"chunk_id": cid,
|
||||
"edge_id": _mk_edge_id(rel, r, cid, "chunk", f"edge_defaults:{note_type}:{rel}"),
|
||||
"provenance": "rule",
|
||||
"rule_id": f"edge_defaults:{note_type}:{rel}",
|
||||
"confidence": 0.7,
|
||||
}))
|
||||
|
||||
refs_all.extend(refs)
|
||||
|
||||
# optional: note-scope references/backlinks
|
||||
# 4) optional note-scope refs/backlinks (+ defaults)
|
||||
if include_note_scope_refs:
|
||||
refs_note = refs_all[:]
|
||||
refs_note = list(refs_all or [])
|
||||
if note_level_references:
|
||||
refs_note.extend([r for r in note_level_references if isinstance(r, str) and r])
|
||||
refs_note = _dedupe(refs_note)
|
||||
refs_note = _dedupe_seq(refs_note)
|
||||
for r in refs_note:
|
||||
edges.append(_edge("references", "note", note_id, r, note_id))
|
||||
edges.append(_edge("backlink", "note", r, note_id, note_id))
|
||||
edges.append(_edge("references", "note", note_id, r, note_id, {
|
||||
"edge_id": _mk_edge_id("references", note_id, r, "note", "explicit:note_scope"),
|
||||
"provenance": "explicit",
|
||||
"rule_id": "explicit:note_scope",
|
||||
"confidence": 1.0,
|
||||
}))
|
||||
edges.append(_edge("backlink", "note", r, note_id, note_id, {
|
||||
"edge_id": _mk_edge_id("backlink", r, note_id, "note", "derived:backlink"),
|
||||
"provenance": "rule",
|
||||
"rule_id": "derived:backlink",
|
||||
"confidence": 0.9,
|
||||
}))
|
||||
for rel in defaults:
|
||||
if rel == "references":
|
||||
continue
|
||||
edges.append(_edge(rel, "note", note_id, r, note_id, {
|
||||
"edge_id": _mk_edge_id(rel, note_id, r, "note", f"edge_defaults:{note_type}:{rel}"),
|
||||
"provenance": "rule",
|
||||
"rule_id": f"edge_defaults:{note_type}:{rel}",
|
||||
"confidence": 0.7,
|
||||
}))
|
||||
if rel in {"related_to", "similar_to"}:
|
||||
edges.append(_edge(rel, "note", r, note_id, note_id, {
|
||||
"edge_id": _mk_edge_id(rel, r, note_id, "note", f"edge_defaults:{note_type}:{rel}"),
|
||||
"provenance": "rule",
|
||||
"rule_id": f"edge_defaults:{note_type}:{rel}",
|
||||
"confidence": 0.7,
|
||||
}))
|
||||
|
||||
return edges
|
||||
# 5) De-Dupe (source_id, target_id, relation, rule_id)
|
||||
seen: Set[Tuple[str,str,str,str]] = set()
|
||||
out: List[dict] = []
|
||||
for e in edges:
|
||||
s = str(e.get("source_id") or "")
|
||||
t = str(e.get("target_id") or "")
|
||||
rel = str(e.get("relation") or e.get("kind") or "edge")
|
||||
rule = str(e.get("rule_id") or "")
|
||||
key = (s, t, rel, rule)
|
||||
if key in seen:
|
||||
continue
|
||||
seen.add(key)
|
||||
out.append(e)
|
||||
return out
|
||||
+257
-81
@@ -2,119 +2,295 @@
|
||||
# -*- coding: utf-8 -*-
|
||||
"""
|
||||
Modul: app/core/edges.py
|
||||
Version: 1.0.0
|
||||
Datum: 2025-09-09
|
||||
Version: 2.0.0 (V2‑superset, rückwärtskompatibel zu v1 vom 2025‑09‑09)
|
||||
|
||||
Zweck
|
||||
-----
|
||||
Zentrale, konsistente Erzeugung von Edge-Payloads im **neuen Schema**:
|
||||
- kind : "belongs_to" | "next" | "prev" | "references" | "backlink"
|
||||
- source_id : ID des Quellknotens (Chunk- oder Note-ID)
|
||||
- target_id : ID des Zielknotens
|
||||
- scope : "chunk" | "note"
|
||||
- note_id : Owner-Note (für performantes Filtern/Löschen)
|
||||
- seq : optional (z. B. Reihenfolge von Vorkommen)
|
||||
Bewahrt die bestehende Edgelogik (belongs_to, prev/next, references, backlink)
|
||||
und ergänzt V2‑Felder + Typ‑Default‑Kanten gemäß config/types.yaml (edge_defaults).
|
||||
Die Funktion ist **idempotent** und **rückwärtskompatibel** zur bisherigen Signatur.
|
||||
|
||||
Hinweise
|
||||
--------
|
||||
- Edges werden dedupliziert (key=(kind,source_id,target_id,scope)).
|
||||
- Für Chunk-Edges wird `note_id` aus dem Chunk-Payload entnommen.
|
||||
- Für Note-Scope-Edges ist `note_id` die Quell-Note-ID.
|
||||
Kompatibilitätsgarantien (gegenüber v1):
|
||||
- **Input**: akzeptiert identische Chunk‑Payloads wie v1:
|
||||
* `id` (Chunk‑ID), `note_id` (Owner), `neighbors.prev|next` (optional),
|
||||
`references: [{target_id: ...}]` (optional),
|
||||
alternativ: `chunk_id`, `chunk_index|ord`, `window|text`
|
||||
- **Output (v1‑Felder)**: `kind`, `source_id`, `target_id`, `scope`, `note_id`, `edge_id`
|
||||
- **Neu (v2‑Felder)**: `relation`, `src_note_id`, `src_chunk_id?`, `dst_note_id`, `dst_chunk_id?`,
|
||||
`provenance` (`explicit|rule`), `rule_id?`, `confidence?`
|
||||
|
||||
Regeln
|
||||
------
|
||||
- Deduplizierungsschlüssel: (source_id, target_id, relation, rule_id)
|
||||
- Strukturkanten:
|
||||
* belongs_to: 1× pro Chunk
|
||||
* next/prev: Sequenz der Chunks; nutzt bevorzugt neighbors; sonst ord/chunk_index
|
||||
- Explizite Referenzen:
|
||||
* aus Chunk: `references[].target_id` (falls vorhanden)
|
||||
* Fallback: Wikilinks in `window|text`: [[Some Title|some-id]] oder [[some-id]]
|
||||
- Note‑Scope:
|
||||
* backlink immer; references nur, wenn include_note_scope_refs=True
|
||||
- Typ‑Defaults (edge_defaults aus config/types.yaml des **Quell‑Notiztyps**):
|
||||
* Für jede explizite Referenz wird je default‑Relation eine Regel‑Kante erzeugt
|
||||
* rule_id: "type_default:{note_type}:{relation}:v1", provenance="rule"
|
||||
|
||||
Konfiguration
|
||||
-------------
|
||||
- ENV MINDNET_TYPES_FILE (Default: ./config/types.yaml)
|
||||
|
||||
Lizenz/Autor
|
||||
------------
|
||||
- Erstimplementierung v1 (2025‑09‑09) — Projekt Mindnet
|
||||
- Erweiterung v2 (2025‑11‑11) — kompatible Superset‑Implementierung
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
from typing import Dict, List
|
||||
|
||||
import os
|
||||
import re
|
||||
from typing import Dict, Iterable, List, Optional, Tuple, Set
|
||||
|
||||
try:
|
||||
import yaml # optional, nur für types.yaml
|
||||
except Exception: # pragma: no cover
|
||||
yaml = None
|
||||
|
||||
# ------------------------------------------------------------
|
||||
# Hilfen: types.yaml laden (edge_defaults)
|
||||
# ------------------------------------------------------------
|
||||
|
||||
def _types_path() -> str:
|
||||
return os.getenv("MINDNET_TYPES_FILE") or "./config/types.yaml"
|
||||
|
||||
def _load_types() -> Dict[str, dict]:
|
||||
p = _types_path()
|
||||
if not os.path.isfile(p) or yaml is None:
|
||||
return {}
|
||||
try:
|
||||
with open(p, "r", encoding="utf-8") as f:
|
||||
data = yaml.safe_load(f) or {}
|
||||
if isinstance(data, dict) and "types" in data and isinstance(data["types"], dict):
|
||||
return data["types"]
|
||||
return data if isinstance(data, dict) else {}
|
||||
except Exception:
|
||||
return {}
|
||||
|
||||
def _edge_defaults_for(note_type: Optional[str]) -> List[str]:
|
||||
types = _load_types()
|
||||
t = (note_type or "").strip().lower()
|
||||
cfg = types.get(t) or {}
|
||||
defaults = cfg.get("edge_defaults") or []
|
||||
if isinstance(defaults, str):
|
||||
defaults = [defaults]
|
||||
return [str(x) for x in defaults if isinstance(x, (str, int, float))]
|
||||
|
||||
# ------------------------------------------------------------
|
||||
# Wikilink‑Parser (Fallback, wenn ch["references"] fehlt)
|
||||
# ------------------------------------------------------------
|
||||
|
||||
_WIKILINK_RE = re.compile(r"\[\[(?:[^\|\]]+\|)?([a-zA-Z0-9_\-#:. ]+)\]\]")
|
||||
|
||||
def _extract_wikilinks(text: str) -> List[str]:
|
||||
ids: List[str] = []
|
||||
for m in _WIKILINK_RE.finditer(text or ""):
|
||||
ids.append(m.group(1).strip())
|
||||
return ids
|
||||
|
||||
# ------------------------------------------------------------
|
||||
# Utility
|
||||
# ------------------------------------------------------------
|
||||
|
||||
def _mk_edge_id(kind: str, s: str, t: str, scope: str, rule_id: Optional[str] = None) -> str:
|
||||
base = f"{kind}:{s}->{t}#{scope}"
|
||||
if rule_id:
|
||||
base += f"|{rule_id}"
|
||||
try:
|
||||
import hashlib
|
||||
return hashlib.blake2s(base.encode("utf-8"), digest_size=12).hexdigest()
|
||||
except Exception: # pragma: no cover
|
||||
return base
|
||||
|
||||
def _dedupe(edges: List[Dict]) -> List[Dict]:
|
||||
seen: Set[Tuple[str,str,str,str]] = set()
|
||||
out: List[Dict] = []
|
||||
for e in edges:
|
||||
s = str(e.get("source_id") or "")
|
||||
t = str(e.get("target_id") or "")
|
||||
rel = str(e.get("relation") or e.get("kind") or "edge")
|
||||
rule = str(e.get("rule_id") or "")
|
||||
key = (s, t, rel, rule)
|
||||
if key in seen:
|
||||
continue
|
||||
seen.add(key)
|
||||
out.append(e)
|
||||
return out
|
||||
|
||||
def _first(v: dict, *keys, default=None):
|
||||
for k in keys:
|
||||
if k in v and v[k] is not None:
|
||||
return v[k]
|
||||
return default
|
||||
|
||||
# ------------------------------------------------------------
|
||||
# Hauptfunktion
|
||||
# ------------------------------------------------------------
|
||||
|
||||
def build_edges_for_note(
|
||||
note_id: str,
|
||||
chunk_payloads: List[Dict],
|
||||
note_level_refs: List[str] | None,
|
||||
note_level_refs: Optional[List[str]] = None,
|
||||
*,
|
||||
include_note_scope_refs: bool = False,
|
||||
) -> List[Dict]:
|
||||
edges: List[Dict] = []
|
||||
chunks = list(chunk_payloads or [])
|
||||
# Notiztyp aus erstem Chunk ableiten (kompatibel zu existierenden Payloads)
|
||||
note_type = (chunks[0].get("type") if chunks else None) or (chunks[0].get("note_type") if chunks else None)
|
||||
|
||||
# Chunk-Scope: belongs_to / prev / next / references
|
||||
for ch in chunk_payloads:
|
||||
cid = ch["id"]
|
||||
# --- Strukturkanten ------------------------------------------------------
|
||||
# belongs_to
|
||||
for ch in chunks:
|
||||
cid = _first(ch, "id", "chunk_id")
|
||||
if not cid:
|
||||
continue
|
||||
owner = ch.get("note_id") or note_id
|
||||
# belongs_to
|
||||
edges.append({
|
||||
e = {
|
||||
"edge_id": _mk_edge_id("belongs_to", cid, note_id, "chunk", "structure:belongs_to:v1"),
|
||||
"kind": "belongs_to",
|
||||
"relation": "belongs_to",
|
||||
"scope": "chunk",
|
||||
"source_id": cid,
|
||||
"target_id": note_id,
|
||||
"scope": "chunk",
|
||||
"note_id": owner,
|
||||
})
|
||||
# Nachbarn
|
||||
"note_id": owner, # v1-Kompat
|
||||
# v2
|
||||
"src_note_id": owner,
|
||||
"src_chunk_id": cid,
|
||||
"dst_note_id": note_id,
|
||||
"provenance": "rule",
|
||||
"rule_id": "structure:belongs_to:v1",
|
||||
"confidence": 1.0,
|
||||
}
|
||||
edges.append(e)
|
||||
|
||||
# next/prev — bevorzugt neighbors.prev/next; sonst via ord/chunk_index
|
||||
# Map der Chunks nach Index
|
||||
ordered = list(chunks)
|
||||
def _idx(c):
|
||||
return _first(c, "chunk_index", "ord", default=0)
|
||||
ordered.sort(key=_idx)
|
||||
|
||||
for i, ch in enumerate(ordered):
|
||||
cid = _first(ch, "id", "chunk_id")
|
||||
if not cid:
|
||||
continue
|
||||
owner = ch.get("note_id") or note_id
|
||||
nb = ch.get("neighbors") or {}
|
||||
prev_id = nb.get("prev")
|
||||
next_id = nb.get("next")
|
||||
# Fallback-Reihenfolge
|
||||
if prev_id is None and i > 0:
|
||||
prev_id = _first(ordered[i-1], "id", "chunk_id")
|
||||
if next_id is None and i+1 < len(ordered):
|
||||
next_id = _first(ordered[i+1], "id", "chunk_id")
|
||||
|
||||
if prev_id:
|
||||
edges.append({
|
||||
"kind": "prev",
|
||||
"source_id": cid,
|
||||
"target_id": prev_id,
|
||||
"scope": "chunk",
|
||||
"note_id": owner,
|
||||
"edge_id": _mk_edge_id("prev", cid, prev_id, "chunk", "structure:order:v1"),
|
||||
"kind": "prev", "relation": "prev", "scope": "chunk",
|
||||
"source_id": cid, "target_id": prev_id, "note_id": owner,
|
||||
"src_note_id": owner, "src_chunk_id": cid,
|
||||
"dst_note_id": owner, "dst_chunk_id": prev_id,
|
||||
"provenance": "rule", "rule_id": "structure:order:v1", "confidence": 0.95,
|
||||
})
|
||||
edges.append({
|
||||
"kind": "next",
|
||||
"source_id": prev_id,
|
||||
"target_id": cid,
|
||||
"scope": "chunk",
|
||||
"note_id": owner,
|
||||
})
|
||||
if next_id:
|
||||
edges.append({
|
||||
"kind": "next",
|
||||
"source_id": cid,
|
||||
"target_id": next_id,
|
||||
"scope": "chunk",
|
||||
"note_id": owner,
|
||||
})
|
||||
edges.append({
|
||||
"kind": "prev",
|
||||
"source_id": next_id,
|
||||
"target_id": cid,
|
||||
"scope": "chunk",
|
||||
"note_id": owner,
|
||||
})
|
||||
# references aus Chunk
|
||||
for ref in (ch.get("references") or []):
|
||||
tid = ref.get("target_id")
|
||||
if not tid:
|
||||
continue
|
||||
edges.append({
|
||||
"kind": "references",
|
||||
"source_id": cid,
|
||||
"target_id": tid,
|
||||
"scope": "chunk",
|
||||
"note_id": owner,
|
||||
"edge_id": _mk_edge_id("next", prev_id, cid, "chunk", "structure:order:v1"),
|
||||
"kind": "next", "relation": "next", "scope": "chunk",
|
||||
"source_id": prev_id, "target_id": cid, "note_id": owner,
|
||||
"src_note_id": owner, "src_chunk_id": prev_id,
|
||||
"dst_note_id": owner, "dst_chunk_id": cid,
|
||||
"provenance": "rule", "rule_id": "structure:order:v1", "confidence": 0.95,
|
||||
})
|
||||
|
||||
# Note-Scope: backlink (immer); references (optional)
|
||||
unique_refs = list(dict.fromkeys(note_level_refs or []))
|
||||
# --- Explizite Referenzen (Chunk‑Scope) ---------------------------------
|
||||
explicit_refs: List[Dict] = []
|
||||
for ch in chunks:
|
||||
cid = _first(ch, "id", "chunk_id")
|
||||
if not cid:
|
||||
continue
|
||||
owner = ch.get("note_id") or note_id
|
||||
# 1) bevorzugt vorhandene ch["references"]
|
||||
refs = ch.get("references") or []
|
||||
targets = [r.get("target_id") for r in refs if isinstance(r, dict) and r.get("target_id")]
|
||||
# 2) Fallback: Wikilinks aus Text
|
||||
if not targets:
|
||||
text = _first(ch, "window", "text", default="") or ""
|
||||
targets = _extract_wikilinks(text)
|
||||
for tid in targets:
|
||||
if not isinstance(tid, str) or not tid.strip():
|
||||
continue
|
||||
e = {
|
||||
"edge_id": _mk_edge_id("references", cid, tid, "chunk"),
|
||||
"kind": "references",
|
||||
"relation": "references",
|
||||
"scope": "chunk",
|
||||
"source_id": cid,
|
||||
"target_id": tid,
|
||||
"note_id": owner,
|
||||
# v2
|
||||
"src_note_id": owner,
|
||||
"src_chunk_id": cid,
|
||||
"dst_note_id": tid,
|
||||
"provenance": "explicit",
|
||||
"rule_id": "",
|
||||
"confidence": 1.0,
|
||||
}
|
||||
edges.append(e)
|
||||
explicit_refs.append(e)
|
||||
|
||||
# --- Note‑Scope: references (optional) + backlink (immer) ----------------
|
||||
unique_refs = []
|
||||
if note_level_refs:
|
||||
seen = set()
|
||||
for tid in note_level_refs:
|
||||
if isinstance(tid, str) and tid.strip() and tid not in seen:
|
||||
unique_refs.append(tid); seen.add(tid)
|
||||
|
||||
for tid in unique_refs:
|
||||
if include_note_scope_refs:
|
||||
edges.append({
|
||||
"kind": "references",
|
||||
"source_id": note_id,
|
||||
"target_id": tid,
|
||||
"scope": "note",
|
||||
"note_id": note_id,
|
||||
"edge_id": _mk_edge_id("references", note_id, tid, "note"),
|
||||
"kind": "references", "relation": "references", "scope": "note",
|
||||
"source_id": note_id, "target_id": tid, "note_id": note_id,
|
||||
"src_note_id": note_id, "dst_note_id": tid,
|
||||
"provenance": "explicit", "rule_id": "", "confidence": 1.0,
|
||||
})
|
||||
edges.append({
|
||||
"kind": "backlink",
|
||||
"source_id": tid,
|
||||
"target_id": note_id,
|
||||
"scope": "note",
|
||||
"note_id": note_id,
|
||||
"edge_id": _mk_edge_id("backlink", tid, note_id, "note", "derived:backlink:v1"),
|
||||
"kind": "backlink", "relation": "backlink", "scope": "note",
|
||||
"source_id": tid, "target_id": note_id, "note_id": note_id,
|
||||
"src_note_id": tid, "dst_note_id": note_id,
|
||||
"provenance": "rule", "rule_id": "derived:backlink:v1", "confidence": 0.9,
|
||||
})
|
||||
|
||||
# Dedupe
|
||||
dedup = {}
|
||||
for e in edges:
|
||||
k = (e["kind"], e["source_id"], e["target_id"], e.get("scope", ""))
|
||||
dedup[k] = e
|
||||
return list(dedup.values())
|
||||
# --- Type‑Defaults je expliziter Referenz --------------------------------
|
||||
defaults = [d for d in _edge_defaults_for(note_type) if d and d != "references"]
|
||||
if defaults:
|
||||
for e in explicit_refs + ([ ] if not include_note_scope_refs else []):
|
||||
# wir nutzen die bereits erzeugten explicit‑Edges als Vorlage
|
||||
src = e["source_id"]; tgt = e["target_id"]
|
||||
scope = e.get("scope", "chunk")
|
||||
s_note = e.get("src_note_id") or note_id
|
||||
s_chunk = e.get("src_chunk_id")
|
||||
t_note = e.get("dst_note_id") or tgt
|
||||
for rel in defaults:
|
||||
rule_id = f"type_default:{(note_type or 'unknown')}:{rel}:v1"
|
||||
edges.append({
|
||||
"edge_id": _mk_edge_id(rel, src, tgt, scope, rule_id),
|
||||
"kind": rel, "relation": rel, "scope": scope,
|
||||
"source_id": src, "target_id": tgt, "note_id": s_note,
|
||||
"src_note_id": s_note, "src_chunk_id": s_chunk,
|
||||
"dst_note_id": t_note,
|
||||
"provenance": "rule", "rule_id": rule_id, "confidence": 0.7,
|
||||
})
|
||||
|
||||
# --- Dedupe & Return -----------------------------------------------------
|
||||
return _dedupe(edges)
|
||||
+210
-46
@@ -1,92 +1,256 @@
|
||||
#!/usr/bin/env python3
|
||||
# -*- coding: utf-8 -*-
|
||||
"""
|
||||
app/core/graph_adapter.py — Adjazenzaufbau & Subgraph-Expansion (WP-04)
|
||||
app/core/graph_adapter.py — Adjazenzaufbau & Subgraph-Expansion
|
||||
|
||||
Zweck:
|
||||
Baut aus Qdrant-Edges (Collection: *_edges) einen leichten In-Memory-Graph
|
||||
und liefert Edge-basierte Kennzahlen (In-Degree, Out-Degree, edge_bonus).
|
||||
Baut aus Qdrant-Edges (Collection: *_edges) einen leichten In-Memory-Graph.
|
||||
|
||||
Kompatibilität:
|
||||
Python 3.12+, qdrant-client 1.x
|
||||
- WP-04a: Liefert Scores (edge_bonus, centrality).
|
||||
- WP-04b: Liefert jetzt auch Struktur-Daten für Erklärungen (Reverse-Lookup).
|
||||
|
||||
Version:
|
||||
0.1.0 (Erstanlage)
|
||||
Stand:
|
||||
2025-10-07
|
||||
Bezug:
|
||||
- WP-04 Edge-Gewichtungen und Heuristiken
|
||||
- app/core/qdrant_points.py (get_edges_for_sources)
|
||||
Nutzung:
|
||||
from app.core.graph_adapter import expand
|
||||
Änderungsverlauf:
|
||||
0.1.0 (2025-10-07) – Erstanlage.
|
||||
0.4.0 (Update für WP-04b: Reverse Adjacency für Explainability)
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
from typing import Dict, List, Optional, DefaultDict
|
||||
from collections import defaultdict
|
||||
from qdrant_client import QdrantClient
|
||||
from app.core.qdrant_points import get_edges_for_sources
|
||||
|
||||
EDGE_BASE_WEIGHTS = {
|
||||
"references": 0.20,
|
||||
"belongs_to": 0.10,
|
||||
"next": 0.06,
|
||||
"prev": 0.06,
|
||||
"backlink": 0.04,
|
||||
"references_at": 0.08,
|
||||
from typing import Dict, List, Optional, DefaultDict, Any
|
||||
from collections import defaultdict
|
||||
|
||||
from qdrant_client import QdrantClient
|
||||
from qdrant_client.http import models as rest
|
||||
|
||||
from app.core.qdrant import collection_names
|
||||
|
||||
# Legacy-Import Fallback
|
||||
try: # pragma: no cover
|
||||
from app.core.qdrant_points import get_edges_for_sources # type: ignore
|
||||
except Exception: # pragma: no cover
|
||||
get_edges_for_sources = None # type: ignore
|
||||
|
||||
|
||||
# Basisgewichte je Edge-Typ (WP-04a Config)
|
||||
EDGE_BASE_WEIGHTS: Dict[str, float] = {
|
||||
# Struktur
|
||||
"belongs_to": 0.10,
|
||||
"next": 0.06,
|
||||
"prev": 0.06,
|
||||
"backlink": 0.04,
|
||||
"references_at": 0.08,
|
||||
|
||||
# Wissen
|
||||
"references": 0.20,
|
||||
"depends_on": 0.18,
|
||||
"related_to": 0.15,
|
||||
"similar_to": 0.12,
|
||||
}
|
||||
|
||||
|
||||
class Subgraph:
|
||||
"""Leichtgewichtiger Subgraph mit Adjazenzlisten & einfachen Kennzahlen."""
|
||||
def _edge_weight(pl: Dict) -> float:
|
||||
"""Berechnet das effektive Edge-Gewicht aus kind + confidence."""
|
||||
kind = pl.get("kind", "edge")
|
||||
base = EDGE_BASE_WEIGHTS.get(kind, 0.0)
|
||||
|
||||
def __init__(self):
|
||||
conf_raw = pl.get("confidence", None)
|
||||
try:
|
||||
conf = float(conf_raw) if conf_raw is not None else None
|
||||
except Exception:
|
||||
conf = None
|
||||
|
||||
if conf is None:
|
||||
return base
|
||||
|
||||
if conf < 0.0: conf = 0.0
|
||||
if conf > 1.0: conf = 1.0
|
||||
|
||||
return base * conf
|
||||
|
||||
|
||||
def _fetch_edges(
|
||||
client: QdrantClient,
|
||||
prefix: str,
|
||||
seeds: List[str],
|
||||
edge_types: Optional[List[str]] = None,
|
||||
limit: int = 2048,
|
||||
) -> List[Dict]:
|
||||
"""
|
||||
Holt Edges direkt aus der *_edges Collection.
|
||||
Filter: source_id IN seeds OR target_id IN seeds OR note_id IN seeds
|
||||
"""
|
||||
if not seeds or limit <= 0:
|
||||
return []
|
||||
|
||||
_, _, edges_col = collection_names(prefix)
|
||||
|
||||
seed_conditions = []
|
||||
for field in ("source_id", "target_id", "note_id"):
|
||||
for s in seeds:
|
||||
seed_conditions.append(
|
||||
rest.FieldCondition(key=field, match=rest.MatchValue(value=str(s)))
|
||||
)
|
||||
seeds_filter = rest.Filter(should=seed_conditions) if seed_conditions else None
|
||||
|
||||
type_filter = None
|
||||
if edge_types:
|
||||
type_conds = [
|
||||
rest.FieldCondition(key="kind", match=rest.MatchValue(value=str(k)))
|
||||
for k in edge_types
|
||||
]
|
||||
type_filter = rest.Filter(should=type_conds)
|
||||
|
||||
must = []
|
||||
if seeds_filter: must.append(seeds_filter)
|
||||
if type_filter: must.append(type_filter)
|
||||
|
||||
flt = rest.Filter(must=must) if must else None
|
||||
|
||||
pts, _ = client.scroll(
|
||||
collection_name=edges_col,
|
||||
scroll_filter=flt,
|
||||
limit=limit,
|
||||
with_payload=True,
|
||||
with_vectors=False,
|
||||
)
|
||||
|
||||
out: List[Dict] = []
|
||||
for p in pts or []:
|
||||
pl = dict(p.payload or {})
|
||||
if pl:
|
||||
out.append(pl)
|
||||
return out
|
||||
|
||||
|
||||
class Subgraph:
|
||||
"""Leichtgewichtiger Subgraph mit Adjazenzlisten & Kennzahlen."""
|
||||
|
||||
def __init__(self) -> None:
|
||||
# Forward: source -> [targets]
|
||||
self.adj: DefaultDict[str, List[Dict]] = defaultdict(list)
|
||||
# Reverse: target -> [sources] (Neu für WP-04b Explanation)
|
||||
self.reverse_adj: DefaultDict[str, List[Dict]] = defaultdict(list)
|
||||
|
||||
self.in_degree: DefaultDict[str, int] = defaultdict(int)
|
||||
self.out_degree: DefaultDict[str, int] = defaultdict(int)
|
||||
|
||||
def add_edge(self, e: Dict):
|
||||
src = e["source"]; tgt = e["target"]; kind = e["kind"]
|
||||
def add_edge(self, e: Dict) -> None:
|
||||
"""
|
||||
Fügt eine Kante hinzu und aktualisiert Forward/Reverse Indizes.
|
||||
e muss enthalten: source, target, kind, weight.
|
||||
"""
|
||||
src = e.get("source")
|
||||
tgt = e.get("target")
|
||||
kind = e.get("kind")
|
||||
weight = e.get("weight", EDGE_BASE_WEIGHTS.get(kind, 0.0))
|
||||
self.adj[src].append({"target": tgt, "kind": kind, "weight": weight})
|
||||
owner = e.get("note_id")
|
||||
|
||||
if not src or not tgt:
|
||||
return
|
||||
|
||||
# 1. Primäre Adjazenz (Forward)
|
||||
edge_data = {"target": tgt, "kind": kind, "weight": weight}
|
||||
self.adj[src].append(edge_data)
|
||||
self.out_degree[src] += 1
|
||||
self.in_degree[tgt] += 1
|
||||
|
||||
# 2. Reverse Adjazenz (Neu für Explanation)
|
||||
# Wir speichern, woher die Kante kam.
|
||||
rev_data = {"source": src, "kind": kind, "weight": weight}
|
||||
self.reverse_adj[tgt].append(rev_data)
|
||||
|
||||
# 3. Kontext-Note Handling (Forward & Reverse)
|
||||
# Wenn eine Kante "im Kontext einer Note" (owner) definiert ist,
|
||||
# schreiben wir sie der Note gut, damit der Retriever Scores auf Note-Ebene findet.
|
||||
if owner and owner != src:
|
||||
# Forward: Owner -> Target
|
||||
self.adj[owner].append(edge_data)
|
||||
self.out_degree[owner] += 1
|
||||
|
||||
# Reverse: Target wird vom Owner referenziert (indirekt)
|
||||
if owner != tgt:
|
||||
rev_owner_data = {"source": owner, "kind": kind, "weight": weight, "via_context": True}
|
||||
self.reverse_adj[tgt].append(rev_owner_data)
|
||||
self.in_degree[owner] += 1 # Leichter Centrality Boost für den Owner
|
||||
|
||||
def aggregate_edge_bonus(self, node_id: str) -> float:
|
||||
"""Summe der ausgehenden Kantengewichte (Hub-Score)."""
|
||||
return sum(edge["weight"] for edge in self.adj.get(node_id, []))
|
||||
|
||||
def edge_bonus(self, node_id: str) -> float:
|
||||
"""API für Retriever (WP-04a Kompatibilität)."""
|
||||
return self.aggregate_edge_bonus(node_id)
|
||||
|
||||
def centrality_bonus(self, node_id: str) -> float:
|
||||
"""Log-gedämpfte Zentralität (In-Degree)."""
|
||||
import math
|
||||
# Log-gedämpfter Bonus, hart begrenzt
|
||||
return min(math.log1p(self.in_degree.get(node_id, 0)) / 10.0, 0.15)
|
||||
indeg = self.in_degree.get(node_id, 0)
|
||||
if indeg <= 0:
|
||||
return 0.0
|
||||
return min(math.log1p(indeg) / 10.0, 0.15)
|
||||
|
||||
# --- WP-04b Explanation Helpers ---
|
||||
|
||||
def get_outgoing_edges(self, node_id: str) -> List[Dict[str, Any]]:
|
||||
"""Liefert Liste aller Ziele, auf die dieser Knoten zeigt."""
|
||||
return self.adj.get(node_id, [])
|
||||
|
||||
def get_incoming_edges(self, node_id: str) -> List[Dict[str, Any]]:
|
||||
"""Liefert Liste aller Quellen, die auf diesen Knoten zeigen."""
|
||||
return self.reverse_adj.get(node_id, [])
|
||||
|
||||
|
||||
def expand(client: QdrantClient, prefix: str, seeds: List[str],
|
||||
depth: int = 1, edge_types: Optional[List[str]] = None) -> Subgraph:
|
||||
def expand(
|
||||
client: QdrantClient,
|
||||
prefix: str,
|
||||
seeds: List[str],
|
||||
depth: int = 1,
|
||||
edge_types: Optional[List[str]] = None,
|
||||
) -> Subgraph:
|
||||
"""
|
||||
Expandiert ab Seeds entlang von Edges (bis depth), optional gefiltert nach Typen.
|
||||
Seeds sind stabile payload-IDs (z. B. note_id, chunk_id).
|
||||
Expandiert ab Seeds entlang von Edges (bis `depth`).
|
||||
"""
|
||||
sg = Subgraph()
|
||||
frontier = set(seeds)
|
||||
visited = set()
|
||||
|
||||
for _ in range(max(depth, 0)):
|
||||
max_depth = max(depth, 0)
|
||||
|
||||
for _ in range(max_depth):
|
||||
if not frontier:
|
||||
break
|
||||
|
||||
edges = get_edges_for_sources(client, prefix, list(frontier), edge_types=edge_types, limit=2048)
|
||||
edges_payloads = _fetch_edges(
|
||||
client=client,
|
||||
prefix=prefix,
|
||||
seeds=list(frontier),
|
||||
edge_types=edge_types,
|
||||
limit=2048,
|
||||
)
|
||||
|
||||
next_frontier = set()
|
||||
for pl in edges:
|
||||
for pl in edges_payloads:
|
||||
src = pl.get("source_id")
|
||||
tgt = pl.get("target_id")
|
||||
|
||||
# Skip invalid edges
|
||||
if not src or not tgt:
|
||||
continue
|
||||
|
||||
e = {
|
||||
"source": pl.get("source_id"),
|
||||
"target": pl.get("target_id"),
|
||||
"source": src,
|
||||
"target": tgt,
|
||||
"kind": pl.get("kind", "edge"),
|
||||
"weight": pl.get("weight", EDGE_BASE_WEIGHTS.get(pl.get("kind", "edge"), 0.0)),
|
||||
"weight": _edge_weight(pl),
|
||||
"note_id": pl.get("note_id"),
|
||||
}
|
||||
sg.add_edge(e)
|
||||
if e["target"]:
|
||||
next_frontier.add(e["target"])
|
||||
|
||||
# Nur weitersuchen, wenn Target noch nicht besucht
|
||||
if tgt and tgt not in visited:
|
||||
next_frontier.add(tgt)
|
||||
|
||||
visited |= frontier
|
||||
frontier = next_frontier - visited
|
||||
|
||||
return sg
|
||||
return sg
|
||||
+179
-57
@@ -1,104 +1,226 @@
|
||||
# note_payload.py
|
||||
#!/usr/bin/env python3
|
||||
# -*- coding: utf-8 -*-
|
||||
"""
|
||||
Modul: app/core/note_payload.py
|
||||
Version: 2.0.0
|
||||
|
||||
Zweck
|
||||
-----
|
||||
Erzeugt ein robustes Note-Payload. Werte wie `retriever_weight`, `chunk_profile`
|
||||
und `edge_defaults` werden in folgender Priorität bestimmt:
|
||||
1) Frontmatter (Note)
|
||||
2) Typ-Registry (config/types.yaml: types.<type>.*)
|
||||
3) Registry-Defaults (config/types.yaml: defaults.*)
|
||||
4) ENV-Defaults (MINDNET_DEFAULT_RETRIEVER_WEIGHT / MINDNET_DEFAULT_CHUNK_PROFILE)
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
from typing import Any, Dict, Optional, Tuple
|
||||
import os, json, pathlib, yaml
|
||||
|
||||
def _as_dict(note: Any) -> Dict[str, Any]:
|
||||
if isinstance(note, dict):
|
||||
return note
|
||||
d: Dict[str, Any] = {}
|
||||
for attr in ("id","note_id","title","path","frontmatter","meta","metadata","type","created","modified","date","tags"):
|
||||
if hasattr(note, attr):
|
||||
d[attr] = getattr(note, attr)
|
||||
# Normalisiere Frontmatter
|
||||
fm = d.get("frontmatter") or d.get("meta") or d.get("metadata") or {}
|
||||
d["frontmatter"] = fm if isinstance(fm, dict) else {}
|
||||
return d
|
||||
from typing import Any, Dict, Tuple, Optional
|
||||
import os
|
||||
import json
|
||||
import pathlib
|
||||
|
||||
def _pick_args(*args, **kwargs) -> Tuple[Optional[str], Optional[Dict[str,Any]]]:
|
||||
path = kwargs.get("path")
|
||||
types_cfg = kwargs.get("types_config")
|
||||
# legacy positional: (path, types_config) oder (types_config, ...)
|
||||
for a in args:
|
||||
if path is None and isinstance(a, (str, pathlib.Path)):
|
||||
path = str(a)
|
||||
if types_cfg is None and isinstance(a, dict):
|
||||
types_cfg = a
|
||||
try:
|
||||
import yaml # type: ignore
|
||||
except Exception:
|
||||
yaml = None
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Helper
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
def _as_dict(x) -> Dict[str, Any]:
|
||||
"""Versucht, ein ParsedMarkdown-ähnliches Objekt in ein Dict zu überführen."""
|
||||
if isinstance(x, dict):
|
||||
return dict(x)
|
||||
|
||||
out: Dict[str, Any] = {}
|
||||
# bekannte Attribute übernehmen, sofern vorhanden
|
||||
for attr in (
|
||||
"frontmatter",
|
||||
"body",
|
||||
"id",
|
||||
"note_id",
|
||||
"title",
|
||||
"path",
|
||||
"tags",
|
||||
"type",
|
||||
"created",
|
||||
"modified",
|
||||
"date",
|
||||
):
|
||||
if hasattr(x, attr):
|
||||
val = getattr(x, attr)
|
||||
if val is not None:
|
||||
out[attr] = val
|
||||
|
||||
# Fallback: wenn immer noch leer, raw speichern
|
||||
if not out:
|
||||
out["raw"] = str(x)
|
||||
|
||||
return out
|
||||
|
||||
|
||||
def _pick_args(*args, **kwargs) -> Tuple[Optional[str], Optional[dict]]:
|
||||
"""Extrahiert optionale Zusatzargumente wie path und types_cfg."""
|
||||
path = kwargs.get("path") or (args[0] if args else None)
|
||||
types_cfg = kwargs.get("types_cfg") or kwargs.get("types") or None
|
||||
return path, types_cfg
|
||||
|
||||
def _load_types_config(explicit: Optional[Dict[str, Any]] = None) -> Dict[str, Any]:
|
||||
if isinstance(explicit, dict):
|
||||
return explicit
|
||||
for rel in ("config/config.yaml", "config/types.yaml"):
|
||||
p = pathlib.Path(rel)
|
||||
if p.exists():
|
||||
with p.open("r", encoding="utf-8") as f:
|
||||
data = yaml.safe_load(f) or {}
|
||||
if isinstance(data, dict) and "types" in data and isinstance(data["types"], dict):
|
||||
return data["types"]
|
||||
return data if isinstance(data, dict) else {}
|
||||
return {}
|
||||
|
||||
def _coalesce(*vals):
|
||||
for v in vals:
|
||||
if v is not None:
|
||||
return v
|
||||
return None
|
||||
|
||||
def _env_float(name: str, default: float) -> float:
|
||||
"""Liest einen Float-Wert aus der Umgebung, mit robustem Fallback."""
|
||||
try:
|
||||
return float(os.environ.get(name, default))
|
||||
except Exception:
|
||||
return default
|
||||
|
||||
|
||||
def _ensure_list(x) -> list:
|
||||
if x is None: return []
|
||||
if isinstance(x, list): return [str(i) for i in x]
|
||||
if isinstance(x, (set, tuple)): return [str(i) for i in x]
|
||||
"""Garantiert eine String-Liste."""
|
||||
if x is None:
|
||||
return []
|
||||
if isinstance(x, list):
|
||||
return [str(i) for i in x]
|
||||
if isinstance(x, (set, tuple)):
|
||||
return [str(i) for i in x]
|
||||
return [str(x)]
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Type-Registry laden
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
def _load_types_config(explicit_cfg: Optional[dict] = None) -> dict:
|
||||
"""Lädt die Type-Registry aus YAML/JSON oder nutzt ein explizit übergebenes Dict."""
|
||||
if explicit_cfg and isinstance(explicit_cfg, dict):
|
||||
return explicit_cfg
|
||||
|
||||
path = os.getenv("MINDNET_TYPES_FILE") or "./config/types.yaml"
|
||||
if not os.path.isfile(path) or yaml is None:
|
||||
return {}
|
||||
|
||||
try:
|
||||
with open(path, "r", encoding="utf-8") as f:
|
||||
data = yaml.safe_load(f) or {}
|
||||
return data if isinstance(data, dict) else {}
|
||||
except Exception:
|
||||
return {}
|
||||
|
||||
|
||||
def _cfg_for_type(note_type: str, reg: dict) -> dict:
|
||||
"""Liefert die Konfiguration für einen konkreten Notiztyp aus der Registry."""
|
||||
if not isinstance(reg, dict):
|
||||
return {}
|
||||
types = reg.get("types") if isinstance(reg.get("types"), dict) else reg
|
||||
return types.get(note_type, {}) if isinstance(types, dict) else {}
|
||||
|
||||
|
||||
def _cfg_defaults(reg: dict) -> dict:
|
||||
"""Liefert den Default-Block aus der Registry (defaults/global)."""
|
||||
if not isinstance(reg, dict):
|
||||
return {}
|
||||
for key in ("defaults", "default", "global"):
|
||||
v = reg.get(key)
|
||||
if isinstance(v, dict):
|
||||
return v
|
||||
return {}
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Haupt-API
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
def make_note_payload(note: Any, *args, **kwargs) -> Dict[str, Any]:
|
||||
"""
|
||||
Baut das Note-Payload für mindnet_notes auf.
|
||||
|
||||
Erwartete Felder im Payload:
|
||||
- note_id: stabile ID aus Frontmatter (id) oder Note-Objekt
|
||||
- title: Titel der Notiz
|
||||
- type: Notiztyp (z. B. concept, project, journal, ...)
|
||||
- path: Pfad im Vault
|
||||
- retriever_weight: effektives Gewicht für den Retriever
|
||||
- chunk_profile: Profil für Chunking (short|medium|long|default|...)
|
||||
- edge_defaults: Liste von Kanten-Typen, die als Defaults gelten
|
||||
"""
|
||||
n = _as_dict(note)
|
||||
path_arg, types_cfg_explicit = _pick_args(*args, **kwargs)
|
||||
types_cfg = _load_types_config(types_cfg_explicit)
|
||||
reg = _load_types_config(types_cfg_explicit)
|
||||
|
||||
fm = n.get("frontmatter") or {}
|
||||
note_type = str(fm.get("type") or n.get("type") or "note")
|
||||
cfg_for_type = types_cfg.get(note_type, {}) if isinstance(types_cfg, dict) else {}
|
||||
fm_type = fm.get("type") or n.get("type") or "concept"
|
||||
note_type = str(fm_type)
|
||||
|
||||
cfg_type = _cfg_for_type(note_type, reg)
|
||||
cfg_def = _cfg_defaults(reg)
|
||||
|
||||
# --- retriever_weight: Frontmatter > Typ-Config > Registry-Defaults > ENV ---
|
||||
default_rw = _env_float("MINDNET_DEFAULT_RETRIEVER_WEIGHT", 1.0)
|
||||
retriever_weight = _coalesce(fm.get("retriever_weight"), cfg_for_type.get("retriever_weight"), default_rw)
|
||||
retriever_weight = fm.get("retriever_weight")
|
||||
if retriever_weight is None:
|
||||
retriever_weight = cfg_type.get(
|
||||
"retriever_weight",
|
||||
cfg_def.get("retriever_weight", default_rw),
|
||||
)
|
||||
try:
|
||||
retriever_weight = float(retriever_weight)
|
||||
except Exception:
|
||||
retriever_weight = default_rw
|
||||
|
||||
chunk_profile = _coalesce(fm.get("chunk_profile"), cfg_for_type.get("chunk_profile"), os.environ.get("MINDNET_DEFAULT_CHUNK_PROFILE","medium"))
|
||||
chunk_profile = chunk_profile if isinstance(chunk_profile, str) else "medium"
|
||||
# --- chunk_profile: Frontmatter > Typ-Config > Registry-Defaults > ENV ---
|
||||
chunk_profile = fm.get("chunk_profile")
|
||||
if chunk_profile is None:
|
||||
chunk_profile = cfg_type.get(
|
||||
"chunk_profile",
|
||||
cfg_def.get(
|
||||
"chunk_profile",
|
||||
os.environ.get("MINDNET_DEFAULT_CHUNK_PROFILE", "medium"),
|
||||
),
|
||||
)
|
||||
if not isinstance(chunk_profile, str):
|
||||
chunk_profile = "medium"
|
||||
|
||||
edge_defaults = _ensure_list(_coalesce(fm.get("edge_defaults"), cfg_for_type.get("edge_defaults"), []))
|
||||
# --- edge_defaults: Frontmatter > Typ-Config > Registry-Defaults ---
|
||||
edge_defaults = fm.get("edge_defaults")
|
||||
if edge_defaults is None:
|
||||
edge_defaults = cfg_type.get(
|
||||
"edge_defaults",
|
||||
cfg_def.get("edge_defaults", []),
|
||||
)
|
||||
edge_defaults = _ensure_list(edge_defaults)
|
||||
|
||||
# --- Basis-Metadaten (IDs, Titel, Pfad) ---
|
||||
note_id = n.get("note_id") or n.get("id") or fm.get("id")
|
||||
title = n.get("title") or fm.get("title") or ""
|
||||
path = n.get("path") or path_arg
|
||||
if isinstance(path, pathlib.Path):
|
||||
path = str(path)
|
||||
|
||||
payload = {
|
||||
payload: Dict[str, Any] = {
|
||||
"note_id": note_id,
|
||||
"title": title,
|
||||
"type": note_type,
|
||||
"path": path or "", # immer vorhanden
|
||||
"path": path or "",
|
||||
"retriever_weight": retriever_weight,
|
||||
"chunk_profile": chunk_profile,
|
||||
"edge_defaults": edge_defaults,
|
||||
}
|
||||
|
||||
# Tags / Keywords übernehmen
|
||||
tags = fm.get("tags") or fm.get("keywords") or n.get("tags")
|
||||
if tags: payload["tags"] = _ensure_list(tags)
|
||||
for k in ("created","modified","date"):
|
||||
v = fm.get(k) or n.get(k)
|
||||
if v: payload[k] = str(v)
|
||||
if tags:
|
||||
payload["tags"] = _ensure_list(tags)
|
||||
|
||||
# Zeitliche Metadaten (sofern vorhanden)
|
||||
for k in ("created", "modified", "date"):
|
||||
v = fm.get(k) or n.get(k)
|
||||
if v:
|
||||
payload[k] = str(v)
|
||||
|
||||
# JSON-Roundtrip zur harten Validierung (ASCII beibehalten)
|
||||
json.loads(json.dumps(payload, ensure_ascii=False))
|
||||
|
||||
return payload
|
||||
+147
-92
@@ -1,124 +1,179 @@
|
||||
#!/usr/bin/env python3
|
||||
# -*- coding: utf-8 -*-
|
||||
"""
|
||||
Name: app/core/qdrant.py
|
||||
Version: v1.4.0 (2025-09-09)
|
||||
app/core/qdrant.py
|
||||
Version: 2.2.0 (2025-11-11)
|
||||
|
||||
Kurzbeschreibung:
|
||||
Qdrant-Client & Collection-Setup für mindnet.
|
||||
- Stellt sicher, dass {prefix}_notes / {prefix}_chunks / {prefix}_edges existieren.
|
||||
- Edges-Collection nutzt 1D Dummy-Vektor.
|
||||
- NEW: ensure_payload_indexes(...) legt sinnvolle Payload-Indizes an.
|
||||
Aufgabe
|
||||
-------
|
||||
- Zentraler Qdrant-Zugriff (Client, Config)
|
||||
- Collection-Anlage (notes/chunks/edges)
|
||||
- **Payload-Indizes sicherstellen** (idempotent)
|
||||
|
||||
Aufruf:
|
||||
from app.core.qdrant import QdrantConfig, get_client, ensure_collections, ensure_payload_indexes
|
||||
Hinweis
|
||||
-------
|
||||
Diese Datei ist als Drop-in-Ersatz gedacht, falls in deinem Projekt noch keine
|
||||
robuste ensure_payload_indexes()-Implementierung vorliegt. Die Signaturen
|
||||
bleiben kompatibel zu scripts.import_markdown und scripts.reset_qdrant.
|
||||
|
||||
API-Notizen
|
||||
-----------
|
||||
- Payload-Indizes werden mit `create_payload_index` angelegt.
|
||||
- Typen stammen aus `qdrant_client.http.models.PayloadSchemaType`:
|
||||
KEYWORD | TEXT | INTEGER | FLOAT | BOOL | GEO | DATETIME
|
||||
- Für häufige Filterfelder (note_id, kind, scope, type, tags, ...) legen wir
|
||||
Indizes an. Das ist laut Qdrant-Doku Best Practice für performante Filter.
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import os
|
||||
from dataclasses import dataclass
|
||||
from typing import Optional, Tuple
|
||||
from typing import Optional, Tuple, Dict, List
|
||||
|
||||
from qdrant_client import QdrantClient
|
||||
from qdrant_client.http import models as rest
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Konfiguration
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
@dataclass
|
||||
class QdrantConfig:
|
||||
url: str
|
||||
api_key: Optional[str]
|
||||
prefix: str
|
||||
dim: int
|
||||
host: Optional[str] = None
|
||||
port: Optional[int] = None
|
||||
url: Optional[str] = None
|
||||
api_key: Optional[str] = None
|
||||
prefix: str = "mindnet"
|
||||
dim: int = 384
|
||||
distance: str = "Cosine" # Cosine | Dot | Euclid
|
||||
on_disk_payload: bool = True
|
||||
|
||||
@staticmethod
|
||||
def from_env() -> "QdrantConfig":
|
||||
url = os.getenv("QDRANT_URL")
|
||||
if not url:
|
||||
host = os.getenv("QDRANT_HOST", "127.0.0.1")
|
||||
port = int(os.getenv("QDRANT_PORT", "6333"))
|
||||
url = f"http://{host}:{port}"
|
||||
@classmethod
|
||||
def from_env(cls) -> "QdrantConfig":
|
||||
# Entweder URL ODER Host/Port, API-Key optional
|
||||
url = os.getenv("QDRANT_URL") or None
|
||||
host = os.getenv("QDRANT_HOST") or None
|
||||
port = os.getenv("QDRANT_PORT")
|
||||
port = int(port) if port else None
|
||||
api_key = os.getenv("QDRANT_API_KEY") or None
|
||||
prefix = os.getenv("COLLECTION_PREFIX", "mindnet")
|
||||
dim = int(os.getenv("VECTOR_DIM", "384"))
|
||||
return QdrantConfig(url=url, api_key=api_key, prefix=prefix, dim=dim)
|
||||
prefix = os.getenv("COLLECTION_PREFIX") or "mindnet"
|
||||
dim = int(os.getenv("VECTOR_DIM") or 384)
|
||||
distance = os.getenv("DISTANCE", "Cosine")
|
||||
on_disk_payload = (os.getenv("ON_DISK_PAYLOAD", "true").lower() == "true")
|
||||
return cls(
|
||||
host=host, port=port, url=url, api_key=api_key,
|
||||
prefix=prefix, dim=dim, distance=distance, on_disk_payload=on_disk_payload
|
||||
)
|
||||
|
||||
|
||||
def get_client(cfg: QdrantConfig) -> QdrantClient:
|
||||
return QdrantClient(url=cfg.url, api_key=cfg.api_key)
|
||||
# QdrantClient akzeptiert entweder url=... oder host/port
|
||||
if cfg.url:
|
||||
return QdrantClient(url=cfg.url, api_key=cfg.api_key, timeout=60.0)
|
||||
return QdrantClient(host=cfg.host or "127.0.0.1", port=cfg.port or 6333, api_key=cfg.api_key, timeout=60.0)
|
||||
|
||||
|
||||
def _create_notes(client: QdrantClient, name: str, dim: int) -> None:
|
||||
if not client.collection_exists(name):
|
||||
client.create_collection(
|
||||
collection_name=name,
|
||||
vectors_config=rest.VectorParams(size=dim, distance=rest.Distance.COSINE),
|
||||
)
|
||||
|
||||
def _create_chunks(client: QdrantClient, name: str, dim: int) -> None:
|
||||
if not client.collection_exists(name):
|
||||
client.create_collection(
|
||||
collection_name=name,
|
||||
vectors_config=rest.VectorParams(size=dim, distance=rest.Distance.COSINE),
|
||||
)
|
||||
|
||||
def _create_edges(client: QdrantClient, name: str) -> None:
|
||||
if not client.collection_exists(name):
|
||||
client.create_collection(
|
||||
collection_name=name,
|
||||
vectors_config=rest.VectorParams(size=1, distance=rest.Distance.DOT), # 1D-Dummy
|
||||
)
|
||||
|
||||
|
||||
def ensure_collections(client: QdrantClient, prefix: str, dim: int, destructive: bool = False) -> None:
|
||||
notes = f"{prefix}_notes"
|
||||
chunks = f"{prefix}_chunks"
|
||||
edges = f"{prefix}_edges"
|
||||
|
||||
_create_notes(client, notes, dim)
|
||||
_create_chunks(client, chunks, dim)
|
||||
|
||||
if client.collection_exists(edges):
|
||||
try:
|
||||
info = client.get_collection(edges)
|
||||
vectors_cfg = getattr(getattr(info.result, "config", None), "params", None)
|
||||
has_vectors = getattr(vectors_cfg, "vectors", None) is not None
|
||||
except Exception:
|
||||
has_vectors = True
|
||||
if not has_vectors:
|
||||
if destructive:
|
||||
client.delete_collection(edges)
|
||||
_create_edges(client, edges)
|
||||
else:
|
||||
print(f"[ensure_collections] WARN: '{edges}' ohne VectorConfig; destructive=False.", flush=True)
|
||||
else:
|
||||
_create_edges(client, edges)
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Collections
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
def collection_names(prefix: str) -> Tuple[str, str, str]:
|
||||
return (f"{prefix}_notes", f"{prefix}_chunks", f"{prefix}_edges")
|
||||
return f"{prefix}_notes", f"{prefix}_chunks", f"{prefix}_edges"
|
||||
|
||||
|
||||
# -------------------------------
|
||||
# NEW: Payload-Indexing
|
||||
# -------------------------------
|
||||
def _vector_params(dim: int, distance: str) -> rest.VectorParams:
|
||||
# Distance: "Cosine" | "Dot" | "Euclid"
|
||||
dist = getattr(rest.Distance, distance.capitalize(), rest.Distance.COSINE)
|
||||
return rest.VectorParams(size=dim, distance=dist)
|
||||
|
||||
def _safe_create_index(client: QdrantClient, col: str, field: str, schema: rest.PayloadSchemaType):
|
||||
try:
|
||||
client.create_payload_index(
|
||||
collection_name=col,
|
||||
field_name=field,
|
||||
field_schema=schema,
|
||||
|
||||
def ensure_collections(client: QdrantClient, prefix: str, dim: int) -> None:
|
||||
"""Legt mindnet_notes, mindnet_chunks, mindnet_edges an (falls nicht vorhanden)."""
|
||||
notes, chunks, edges = collection_names(prefix)
|
||||
|
||||
# notes
|
||||
if not client.collection_exists(notes):
|
||||
client.create_collection(
|
||||
collection_name=notes,
|
||||
vectors_config=_vector_params(dim, os.getenv("DISTANCE", "Cosine")),
|
||||
on_disk_payload=True,
|
||||
)
|
||||
except Exception:
|
||||
# bereits vorhanden oder nicht unterstütztes Schema → ignorieren
|
||||
pass
|
||||
# chunks
|
||||
if not client.collection_exists(chunks):
|
||||
client.create_collection(
|
||||
collection_name=chunks,
|
||||
vectors_config=_vector_params(dim, os.getenv("DISTANCE", "Cosine")),
|
||||
on_disk_payload=True,
|
||||
)
|
||||
# edges (Dummy-Vektor, Filter via Payload)
|
||||
if not client.collection_exists(edges):
|
||||
client.create_collection(
|
||||
collection_name=edges,
|
||||
vectors_config=_vector_params(1, "Dot"),
|
||||
on_disk_payload=True,
|
||||
)
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Payload-Indizes
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
def _ensure_index(client: QdrantClient, collection: str, field: str, schema: rest.PayloadSchemaType) -> None:
|
||||
"""Idempotentes Anlegen eines Payload-Indexes für ein Feld."""
|
||||
try:
|
||||
client.create_payload_index(collection_name=collection, field_name=field, field_schema=schema, wait=True)
|
||||
except Exception as e:
|
||||
# Fehler ignorieren, falls Index bereits existiert oder Server "already indexed" meldet.
|
||||
# Für Debugging ggf. Logging ergänzen.
|
||||
_ = e
|
||||
|
||||
|
||||
def ensure_payload_indexes(client: QdrantClient, prefix: str) -> None:
|
||||
"""
|
||||
Stellt sicher, dass alle benötigten Payload-Indizes existieren.
|
||||
- notes: note_id(KEYWORD), type(KEYWORD), title(TEXT), updated(INTEGER), tags(KEYWORD)
|
||||
- chunks: note_id(KEYWORD), chunk_id(KEYWORD), index(INTEGER), type(KEYWORD), tags(KEYWORD)
|
||||
- edges: note_id(KEYWORD), kind(KEYWORD), scope(KEYWORD), source_id(KEYWORD), target_id(KEYWORD), chunk_id(KEYWORD)
|
||||
"""
|
||||
notes, chunks, edges = collection_names(prefix)
|
||||
# Notes
|
||||
_safe_create_index(client, notes, "note_id", rest.PayloadSchemaType.KEYWORD)
|
||||
# Chunks
|
||||
_safe_create_index(client, chunks, "note_id", rest.PayloadSchemaType.KEYWORD)
|
||||
_safe_create_index(client, chunks, "chunk_index", rest.PayloadSchemaType.INTEGER)
|
||||
# Edges
|
||||
for f in ("kind", "scope", "source_id", "target_id", "note_id"):
|
||||
_safe_create_index(client, edges, f, rest.PayloadSchemaType.KEYWORD)
|
||||
|
||||
# NOTES
|
||||
for field, schema in [
|
||||
("note_id", rest.PayloadSchemaType.KEYWORD),
|
||||
("type", rest.PayloadSchemaType.KEYWORD),
|
||||
("title", rest.PayloadSchemaType.TEXT),
|
||||
("updated", rest.PayloadSchemaType.INTEGER),
|
||||
("tags", rest.PayloadSchemaType.KEYWORD),
|
||||
]:
|
||||
_ensure_index(client, notes, field, schema)
|
||||
|
||||
# CHUNKS
|
||||
for field, schema in [
|
||||
("note_id", rest.PayloadSchemaType.KEYWORD),
|
||||
("chunk_id", rest.PayloadSchemaType.KEYWORD),
|
||||
("index", rest.PayloadSchemaType.INTEGER),
|
||||
("type", rest.PayloadSchemaType.KEYWORD),
|
||||
("tags", rest.PayloadSchemaType.KEYWORD),
|
||||
]:
|
||||
_ensure_index(client, chunks, field, schema)
|
||||
|
||||
# EDGES
|
||||
for field, schema in [
|
||||
("note_id", rest.PayloadSchemaType.KEYWORD),
|
||||
("kind", rest.PayloadSchemaType.KEYWORD),
|
||||
("scope", rest.PayloadSchemaType.KEYWORD),
|
||||
("source_id", rest.PayloadSchemaType.KEYWORD),
|
||||
("target_id", rest.PayloadSchemaType.KEYWORD),
|
||||
("chunk_id", rest.PayloadSchemaType.KEYWORD),
|
||||
]:
|
||||
_ensure_index(client, edges, field, schema)
|
||||
|
||||
|
||||
__all__ = [
|
||||
"QdrantConfig",
|
||||
"get_client",
|
||||
"ensure_collections",
|
||||
"ensure_payload_indexes",
|
||||
"collection_names",
|
||||
]
|
||||
@@ -217,3 +217,84 @@ def search_chunks_by_vector(client: QdrantClient, prefix: str, vector: List[floa
|
||||
for r in res:
|
||||
out.append((str(r.id), float(r.score), dict(r.payload or {})))
|
||||
return out
|
||||
|
||||
|
||||
# --- Edge retrieval helper ---
|
||||
|
||||
def get_edges_for_sources(
|
||||
client: QdrantClient,
|
||||
prefix: str,
|
||||
source_ids: Iterable[str],
|
||||
edge_types: Optional[Iterable[str]] = None,
|
||||
limit: int = 2048,
|
||||
) -> List[Dict[str, Any]]:
|
||||
"""Retrieve edge payloads from the <prefix>_edges collection.
|
||||
|
||||
Args:
|
||||
client: QdrantClient instance.
|
||||
prefix: Mindnet collection prefix (e.g. "mindnet").
|
||||
source_ids: Iterable of source_id values (typically chunk_ids or note_ids).
|
||||
edge_types: Optional iterable of edge kinds (e.g. ["references", "depends_on"]). If None,
|
||||
all kinds are returned.
|
||||
limit: Maximum number of edge payloads to return.
|
||||
|
||||
Returns:
|
||||
A list of edge payload dicts, e.g.:
|
||||
{
|
||||
"note_id": "...",
|
||||
"chunk_id": "...",
|
||||
"kind": "references" | "depends_on" | ...,
|
||||
"scope": "chunk",
|
||||
"source_id": "...",
|
||||
"target_id": "...",
|
||||
"rule_id": "...",
|
||||
"confidence": 0.7,
|
||||
...
|
||||
}
|
||||
"""
|
||||
source_ids = list(source_ids)
|
||||
if not source_ids or limit <= 0:
|
||||
return []
|
||||
|
||||
# Resolve collection name
|
||||
_, _, edges_col = _names(prefix)
|
||||
|
||||
# Build filter: source_id IN source_ids
|
||||
src_filter = _filter_any("source_id", [str(s) for s in source_ids])
|
||||
|
||||
# Optional: kind IN edge_types
|
||||
kind_filter = None
|
||||
if edge_types:
|
||||
kind_filter = _filter_any("kind", [str(k) for k in edge_types])
|
||||
|
||||
flt = _merge_filters(src_filter, kind_filter)
|
||||
|
||||
out: List[Dict[str, Any]] = []
|
||||
next_page = None
|
||||
remaining = int(limit)
|
||||
|
||||
# Use paginated scroll API; we don't need vectors, only payloads.
|
||||
while remaining > 0:
|
||||
batch_limit = min(256, remaining)
|
||||
res, next_page = client.scroll(
|
||||
collection_name=edges_col,
|
||||
scroll_filter=flt,
|
||||
limit=batch_limit,
|
||||
with_payload=True,
|
||||
with_vectors=False,
|
||||
offset=next_page,
|
||||
)
|
||||
|
||||
if not res:
|
||||
break
|
||||
|
||||
for r in res:
|
||||
out.append(dict(r.payload or {}))
|
||||
remaining -= 1
|
||||
if remaining <= 0:
|
||||
break
|
||||
|
||||
if next_page is None or remaining <= 0:
|
||||
break
|
||||
|
||||
return out
|
||||
+389
-99
@@ -1,121 +1,411 @@
|
||||
"""
|
||||
app/core/retriever.py — Semantischer/Edge-Aware/Hybrid Retriever (WP-04)
|
||||
|
||||
Zweck:
|
||||
Kandidatenfindung via Vektorsuche in *_chunks, optionale Edge-Expansion
|
||||
und kombiniertes Ranking zur Rückgabe von Top-K Treffern.
|
||||
Erweiterung (0.2.0): Text→Embedding, falls kein query_vector übergeben wurde.
|
||||
|
||||
Kompatibilität:
|
||||
Python 3.12+, qdrant-client 1.x
|
||||
Version:
|
||||
0.2.0 (Text→Embedding ergänzt; bestehendes Verhalten unverändert)
|
||||
Stand:
|
||||
2025-10-07
|
||||
Bezug:
|
||||
- app/core/graph_adapter.py (expand)
|
||||
- app/core/ranking.py (combine_scores)
|
||||
- app/core/qdrant_points.py (search_chunks_by_vector)
|
||||
- app/services/embeddings_client.py (embed_text)
|
||||
- app/models/dto.py (QueryRequest/Response)
|
||||
Änderungsverlauf:
|
||||
0.2.0 (2025-10-07) – Text→Embedding (embed_text_if_needed).
|
||||
0.1.0 (2025-10-07) – Erstanlage.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import os
|
||||
import time
|
||||
from typing import Dict, List, Optional, Tuple
|
||||
from qdrant_client import QdrantClient
|
||||
from functools import lru_cache
|
||||
from typing import Any, Dict, List, Tuple, Iterable, Optional
|
||||
|
||||
from app.models.dto import QueryRequest, QueryResponse, QueryHit
|
||||
from app.core.ranking import combine_scores
|
||||
from app.core.graph_adapter import expand
|
||||
from app.core import qdrant_points as qp
|
||||
from app.config import get_settings
|
||||
from app.services.embeddings_client import embed_text
|
||||
from app.models.dto import (
|
||||
QueryRequest,
|
||||
QueryResponse,
|
||||
QueryHit,
|
||||
Explanation,
|
||||
ScoreBreakdown,
|
||||
Reason,
|
||||
EdgeDTO
|
||||
)
|
||||
import app.core.qdrant as qdr
|
||||
import app.core.qdrant_points as qp
|
||||
import app.services.embeddings_client as ec
|
||||
import app.core.graph_adapter as ga
|
||||
|
||||
try:
|
||||
import yaml # type: ignore[import]
|
||||
except Exception: # pragma: no cover - Fallback, falls PyYAML nicht installiert ist
|
||||
yaml = None # type: ignore[assignment]
|
||||
|
||||
|
||||
def _vector_from_request(req: QueryRequest) -> List[float]:
|
||||
@lru_cache
|
||||
def _get_scoring_weights() -> Tuple[float, float, float]:
|
||||
"""Liefert (semantic_weight, edge_weight, centrality_weight) für den Retriever.
|
||||
|
||||
Priorität:
|
||||
1. Werte aus config/retriever.yaml (falls vorhanden und gültig).
|
||||
2. Fallback auf Settings.RETRIEVER_W_* (ENV-basiert).
|
||||
"""
|
||||
Query-Vektor bestimmen:
|
||||
- Falls query_vector gesetzt: unverändert verwenden (Back-compat, Tests).
|
||||
- Sonst, falls query gesetzt: serverseitig einbetten.
|
||||
- Andernfalls: Fehler.
|
||||
settings = get_settings()
|
||||
sem = float(getattr(settings, "RETRIEVER_W_SEM", 1.0))
|
||||
edge = float(getattr(settings, "RETRIEVER_W_EDGE", 0.0))
|
||||
cent = float(getattr(settings, "RETRIEVER_W_CENT", 0.0))
|
||||
|
||||
# YAML-Override, falls konfiguriert
|
||||
config_path = os.getenv("MINDNET_RETRIEVER_CONFIG", "config/retriever.yaml")
|
||||
if yaml is None:
|
||||
return sem, edge, cent
|
||||
try:
|
||||
if os.path.exists(config_path):
|
||||
with open(config_path, "r", encoding="utf-8") as f:
|
||||
data = yaml.safe_load(f) or {}
|
||||
scoring = data.get("scoring", {}) or {}
|
||||
sem = float(scoring.get("semantic_weight", sem))
|
||||
edge = float(scoring.get("edge_weight", edge))
|
||||
cent = float(scoring.get("centrality_weight", cent))
|
||||
except Exception:
|
||||
# Bei Fehlern in der YAML-Konfiguration defensiv auf Defaults zurückfallen
|
||||
return sem, edge, cent
|
||||
return sem, edge, cent
|
||||
|
||||
|
||||
def _get_client_and_prefix() -> Tuple[Any, str]:
|
||||
"""Liefert (QdrantClient, prefix) basierend auf QdrantConfig.from_env()."""
|
||||
cfg = qdr.QdrantConfig.from_env()
|
||||
client = qdr.get_client(cfg)
|
||||
return client, cfg.prefix
|
||||
|
||||
|
||||
def _get_query_vector(req: QueryRequest) -> List[float]:
|
||||
"""
|
||||
Liefert den Query-Vektor aus dem Request.
|
||||
"""
|
||||
if req.query_vector:
|
||||
return req.query_vector
|
||||
if req.query:
|
||||
return embed_text(req.query)
|
||||
raise ValueError("query_vector fehlt. Alternativ 'query' (Text) übergeben, wird serverseitig eingebettet.")
|
||||
return list(req.query_vector)
|
||||
|
||||
if not req.query:
|
||||
raise ValueError("QueryRequest benötigt entweder query oder query_vector")
|
||||
|
||||
settings = get_settings()
|
||||
model_name = settings.MODEL_NAME
|
||||
|
||||
try:
|
||||
return ec.embed_text(req.query, model_name=model_name) # type: ignore[call-arg]
|
||||
except TypeError:
|
||||
return ec.embed_text(req.query) # type: ignore[call-arg]
|
||||
|
||||
|
||||
def _semantic_hits(
|
||||
client: Any,
|
||||
prefix: str,
|
||||
vector: List[float],
|
||||
top_k: int,
|
||||
filters: Dict[str, Any] | None = None,
|
||||
) -> List[Tuple[str, float, Dict[str, Any]]]:
|
||||
"""Führt eine semantische Suche über mindnet_chunks aus und liefert Roh-Treffer."""
|
||||
flt = filters or None
|
||||
raw_hits = qp.search_chunks_by_vector(client, prefix, vector, top=top_k, filters=flt)
|
||||
results: List[Tuple[str, float, Dict[str, Any]]] = []
|
||||
for pid, score, payload in raw_hits:
|
||||
results.append((str(pid), float(score), dict(payload or {})))
|
||||
return results
|
||||
|
||||
|
||||
def _compute_total_score(
|
||||
semantic_score: float,
|
||||
payload: Dict[str, Any],
|
||||
edge_bonus: float = 0.0,
|
||||
cent_bonus: float = 0.0,
|
||||
) -> Tuple[float, float, float]:
|
||||
"""Berechnet total_score aus semantic_score, retriever_weight und Graph-Boni."""
|
||||
raw_weight = payload.get("retriever_weight", 1.0)
|
||||
try:
|
||||
weight = float(raw_weight)
|
||||
except (TypeError, ValueError):
|
||||
weight = 1.0
|
||||
if weight < 0.0:
|
||||
weight = 0.0
|
||||
|
||||
sem_w, edge_w, cent_w = _get_scoring_weights()
|
||||
total = (sem_w * float(semantic_score) * weight) + (edge_w * edge_bonus) + (cent_w * cent_bonus)
|
||||
return float(total), float(edge_bonus), float(cent_bonus)
|
||||
|
||||
|
||||
# --- WP-04b Explanation Logic ---
|
||||
|
||||
def _build_explanation(
|
||||
semantic_score: float,
|
||||
payload: Dict[str, Any],
|
||||
edge_bonus: float,
|
||||
cent_bonus: float,
|
||||
subgraph: Optional[ga.Subgraph],
|
||||
node_key: Optional[str]
|
||||
) -> Explanation:
|
||||
"""
|
||||
Erstellt ein detailliertes Explanation-Objekt für einen Treffer.
|
||||
Analysiert Scores, Typen und Kanten (Incoming & Outgoing).
|
||||
"""
|
||||
sem_w, edge_w, cent_w = _get_scoring_weights()
|
||||
|
||||
try:
|
||||
type_weight = float(payload.get("retriever_weight", 1.0))
|
||||
except (TypeError, ValueError):
|
||||
type_weight = 1.0
|
||||
|
||||
note_type = payload.get("type", "unknown")
|
||||
|
||||
breakdown = ScoreBreakdown(
|
||||
semantic_contribution=(sem_w * semantic_score * type_weight),
|
||||
edge_contribution=(edge_w * edge_bonus),
|
||||
centrality_contribution=(cent_w * cent_bonus),
|
||||
raw_semantic=semantic_score,
|
||||
raw_edge_bonus=edge_bonus,
|
||||
raw_centrality=cent_bonus,
|
||||
node_weight=type_weight
|
||||
)
|
||||
|
||||
reasons: List[Reason] = []
|
||||
edges_dto: List[EdgeDTO] = []
|
||||
|
||||
# 1. Semantische Gründe
|
||||
if semantic_score > 0.85:
|
||||
reasons.append(Reason(
|
||||
kind="semantic",
|
||||
message="Sehr hohe textuelle Übereinstimmung.",
|
||||
score_impact=breakdown.semantic_contribution
|
||||
))
|
||||
elif semantic_score > 0.70:
|
||||
reasons.append(Reason(
|
||||
kind="semantic",
|
||||
message="Gute textuelle Übereinstimmung.",
|
||||
score_impact=breakdown.semantic_contribution
|
||||
))
|
||||
|
||||
# 2. Typ-Gründe
|
||||
if type_weight != 1.0:
|
||||
msg = "Bevorzugt" if type_weight > 1.0 else "Leicht abgewertet"
|
||||
reasons.append(Reason(
|
||||
kind="type",
|
||||
message=f"{msg} aufgrund des Typs '{note_type}' (Gewicht: {type_weight}).",
|
||||
score_impact=(sem_w * semantic_score * (type_weight - 1.0))
|
||||
))
|
||||
|
||||
# 3. Graph-Gründe (Edges)
|
||||
if subgraph and node_key and edge_bonus > 0:
|
||||
# Wir sammeln die stärksten Kanten (egal ob rein oder raus),
|
||||
# die zum Score beitragen.
|
||||
|
||||
# A) Outgoing (Ich verweise auf...) - Das ist oft der Hub-Score
|
||||
if hasattr(subgraph, "get_outgoing_edges"):
|
||||
outgoing = subgraph.get_outgoing_edges(node_key)
|
||||
for edge in outgoing:
|
||||
target = edge.get("target", "Unknown")
|
||||
kind = edge.get("kind", "edge")
|
||||
weight = edge.get("weight", 0.0)
|
||||
|
||||
# Nur relevante Kanten aufnehmen
|
||||
if weight > 0.05:
|
||||
edges_dto.append(EdgeDTO(
|
||||
id=f"{node_key}->{target}:{kind}",
|
||||
kind=kind, source=node_key, target=target, weight=weight, direction="out"
|
||||
))
|
||||
|
||||
# B) Incoming (Ich werde verwiesen von...)
|
||||
if hasattr(subgraph, "get_incoming_edges"):
|
||||
incoming = subgraph.get_incoming_edges(node_key)
|
||||
for edge in incoming:
|
||||
src = edge.get("source", "Unknown")
|
||||
kind = edge.get("kind", "edge")
|
||||
weight = edge.get("weight", 0.0)
|
||||
|
||||
if weight > 0.05:
|
||||
edges_dto.append(EdgeDTO(
|
||||
id=f"{src}->{node_key}:{kind}",
|
||||
kind=kind, source=src, target=node_key, weight=weight, direction="in"
|
||||
))
|
||||
|
||||
# Sortieren nach Gewicht und Top-3 als Reasons generieren
|
||||
all_edges = sorted(edges_dto, key=lambda e: e.weight, reverse=True)
|
||||
|
||||
for top_edge in all_edges[:3]:
|
||||
# Impact schätzen (grob, da Edge-Bonus eine Summe ist)
|
||||
impact = edge_w * top_edge.weight
|
||||
|
||||
if top_edge.direction == "out":
|
||||
msg = f"Verweist auf '{top_edge.target}' via '{top_edge.kind}'"
|
||||
else:
|
||||
msg = f"Referenziert von '{top_edge.source}' via '{top_edge.kind}'"
|
||||
|
||||
reasons.append(Reason(
|
||||
kind="edge",
|
||||
message=msg,
|
||||
score_impact=impact,
|
||||
details={"kind": top_edge.kind, "weight": top_edge.weight}
|
||||
))
|
||||
|
||||
# 4. Centrality
|
||||
if cent_bonus > 0.01:
|
||||
reasons.append(Reason(
|
||||
kind="centrality",
|
||||
message="Knoten liegt zentral im Kontext.",
|
||||
score_impact=breakdown.centrality_contribution
|
||||
))
|
||||
|
||||
return Explanation(
|
||||
breakdown=breakdown,
|
||||
reasons=reasons,
|
||||
related_edges=edges_dto if edges_dto else None
|
||||
)
|
||||
# --- End Explanation Logic ---
|
||||
|
||||
|
||||
def _extract_expand_options(req: QueryRequest) -> Tuple[int, List[str] | None]:
|
||||
"""Extrahiert depth und edge_types aus req.expand."""
|
||||
expand = getattr(req, "expand", None)
|
||||
if not expand:
|
||||
return 0, None
|
||||
|
||||
depth = 1
|
||||
edge_types: List[str] | None = None
|
||||
|
||||
if hasattr(expand, "depth") or hasattr(expand, "edge_types"):
|
||||
depth = int(getattr(expand, "depth", 1) or 1)
|
||||
types_val = getattr(expand, "edge_types", None)
|
||||
if types_val:
|
||||
edge_types = list(types_val)
|
||||
return depth, edge_types
|
||||
|
||||
if isinstance(expand, dict):
|
||||
if "depth" in expand:
|
||||
depth = int(expand.get("depth") or 1)
|
||||
if "edge_types" in expand and expand["edge_types"] is not None:
|
||||
edge_types = list(expand["edge_types"])
|
||||
return depth, edge_types
|
||||
|
||||
return 0, None
|
||||
|
||||
|
||||
def _build_hits_from_semantic(
|
||||
hits: Iterable[Tuple[str, float, Dict[str, Any]]],
|
||||
top_k: int,
|
||||
used_mode: str,
|
||||
subgraph: ga.Subgraph | None = None,
|
||||
explain: bool = False, # WP-04b
|
||||
) -> QueryResponse:
|
||||
"""Baut aus Raw-Hits und optionalem Subgraph strukturierte QueryHits.
|
||||
|
||||
WP-04b: Wenn explain=True, wird _build_explanation aufgerufen.
|
||||
"""
|
||||
t0 = time.time()
|
||||
|
||||
enriched: List[Tuple[str, float, Dict[str, Any], float, float, float]] = []
|
||||
|
||||
for pid, semantic_score, payload in hits:
|
||||
edge_bonus = 0.0
|
||||
cent_bonus = 0.0
|
||||
node_key = payload.get("chunk_id") or payload.get("note_id")
|
||||
|
||||
if subgraph is not None and node_key:
|
||||
try:
|
||||
edge_bonus = float(subgraph.edge_bonus(node_key))
|
||||
except Exception:
|
||||
edge_bonus = 0.0
|
||||
try:
|
||||
cent_bonus = float(subgraph.centrality_bonus(node_key))
|
||||
except Exception:
|
||||
cent_bonus = 0.0
|
||||
|
||||
total, edge_bonus, cent_bonus = _compute_total_score(
|
||||
semantic_score,
|
||||
payload,
|
||||
edge_bonus=edge_bonus,
|
||||
cent_bonus=cent_bonus,
|
||||
)
|
||||
enriched.append((pid, float(semantic_score), payload, total, edge_bonus, cent_bonus))
|
||||
|
||||
# Sortierung nach total_score absteigend
|
||||
enriched_sorted = sorted(enriched, key=lambda h: h[3], reverse=True)
|
||||
limited = enriched_sorted[: max(1, top_k)]
|
||||
|
||||
results: List[QueryHit] = []
|
||||
for pid, semantic_score, payload, total, edge_bonus, cent_bonus in limited:
|
||||
note_id = payload.get("note_id")
|
||||
path = payload.get("path")
|
||||
section = payload.get("section") or payload.get("section_title")
|
||||
node_key = payload.get("chunk_id") or payload.get("note_id")
|
||||
|
||||
# WP-04b: Explanation bauen?
|
||||
explanation_obj = None
|
||||
if explain:
|
||||
explanation_obj = _build_explanation(
|
||||
semantic_score=float(semantic_score),
|
||||
payload=payload,
|
||||
edge_bonus=edge_bonus,
|
||||
cent_bonus=cent_bonus,
|
||||
subgraph=subgraph,
|
||||
node_key=node_key
|
||||
)
|
||||
|
||||
results.append(
|
||||
QueryHit(
|
||||
node_id=str(pid),
|
||||
note_id=note_id,
|
||||
semantic_score=float(semantic_score),
|
||||
edge_bonus=edge_bonus,
|
||||
centrality_bonus=cent_bonus,
|
||||
total_score=total,
|
||||
paths=None,
|
||||
source={
|
||||
"path": path,
|
||||
"section": section,
|
||||
},
|
||||
explanation=explanation_obj
|
||||
)
|
||||
)
|
||||
|
||||
dt = int((time.time() - t0) * 1000)
|
||||
return QueryResponse(results=results, used_mode=used_mode, latency_ms=dt)
|
||||
|
||||
|
||||
def semantic_retrieve(req: QueryRequest) -> QueryResponse:
|
||||
"""Nur semantische Kandidaten, keine Edge-Expansion (depth=0)."""
|
||||
t0 = time.time()
|
||||
s = get_settings()
|
||||
client = QdrantClient(url=s.QDRANT_URL, api_key=s.QDRANT_API_KEY)
|
||||
"""Reiner semantischer Retriever (ohne Edge-Expansion)."""
|
||||
client, prefix = _get_client_and_prefix()
|
||||
vector = _get_query_vector(req)
|
||||
top_k = req.top_k or get_settings().RETRIEVER_TOP_K
|
||||
|
||||
q_vec = _vector_from_request(req)
|
||||
raw_hits = qp.search_chunks_by_vector(client, s.COLLECTION_PREFIX, q_vec, top=req.top_k, filters=req.filters)
|
||||
|
||||
results: List[QueryHit] = []
|
||||
for pid, s_score, payload in raw_hits:
|
||||
results.append(QueryHit(
|
||||
node_id=pid,
|
||||
note_id=payload.get("note_id"),
|
||||
semantic_score=float(s_score),
|
||||
edge_bonus=0.0,
|
||||
centrality_bonus=0.0,
|
||||
total_score=float(s_score), # un-normalisiert: ok für quick semantic mode
|
||||
paths=None,
|
||||
source={"path": payload.get("path"), "section": payload.get("section_title")}
|
||||
))
|
||||
dt = int((time.time() - t0) * 1000)
|
||||
return QueryResponse(results=results, used_mode="semantic", latency_ms=dt)
|
||||
hits = _semantic_hits(client, prefix, vector, top_k=top_k, filters=req.filters)
|
||||
|
||||
# explain Flag durchreichen
|
||||
return _build_hits_from_semantic(
|
||||
hits,
|
||||
top_k=top_k,
|
||||
used_mode="semantic",
|
||||
subgraph=None,
|
||||
explain=req.explain
|
||||
)
|
||||
|
||||
|
||||
def hybrid_retrieve(req: QueryRequest) -> QueryResponse:
|
||||
"""Semantik + Edge-Expansion + kombiniertes Ranking."""
|
||||
t0 = time.time()
|
||||
s = get_settings()
|
||||
client = QdrantClient(url=s.QDRANT_URL, api_key=s.QDRANT_API_KEY)
|
||||
"""Hybrid-Retriever: semantische Suche + optionale Edge-Expansion."""
|
||||
client, prefix = _get_client_and_prefix()
|
||||
|
||||
q_vec = _vector_from_request(req)
|
||||
if req.query_vector:
|
||||
vector = list(req.query_vector)
|
||||
else:
|
||||
vector = _get_query_vector(req)
|
||||
|
||||
# 1) Semantische Seeds (top_k * 3 für breitere Basis)
|
||||
raw_hits = qp.search_chunks_by_vector(client, s.COLLECTION_PREFIX, q_vec, top=req.top_k * 3, filters=req.filters)
|
||||
id2payload = {pid: payload for (pid, _, payload) in raw_hits}
|
||||
seeds = [pid for (pid, _, _) in raw_hits]
|
||||
top_k = req.top_k or get_settings().RETRIEVER_TOP_K
|
||||
|
||||
# 2) Edge-Expansion
|
||||
edge_types = req.expand.get("edge_types") if req.expand else None
|
||||
depth = req.expand.get("depth", 1) if req.expand else 1
|
||||
sg = expand(client, s.COLLECTION_PREFIX, seeds, depth=depth, edge_types=edge_types)
|
||||
hits = _semantic_hits(client, prefix, vector, top_k=top_k, filters=req.filters)
|
||||
|
||||
edge_bonus_map = {pid: sg.aggregate_edge_bonus(pid) for pid in seeds}
|
||||
centrality_map = {pid: sg.centrality_bonus(pid) for pid in seeds}
|
||||
depth, edge_types = _extract_expand_options(req)
|
||||
subgraph: ga.Subgraph | None = None
|
||||
if depth and depth > 0:
|
||||
seed_ids: List[str] = []
|
||||
for _pid, _score, payload in hits:
|
||||
key = payload.get("chunk_id") or payload.get("note_id")
|
||||
if key and key not in seed_ids:
|
||||
seed_ids.append(key)
|
||||
|
||||
# 3) Combined Ranking
|
||||
scored = combine_scores(
|
||||
raw_hits, edge_bonus_map, centrality_map,
|
||||
w_sem=s.RETRIEVER_W_SEM, w_edge=s.RETRIEVER_W_EDGE, w_cent=s.RETRIEVER_W_CENT
|
||||
)
|
||||
if seed_ids:
|
||||
try:
|
||||
subgraph = ga.expand(client, prefix, seed_ids, depth=depth, edge_types=edge_types)
|
||||
except Exception:
|
||||
subgraph = None
|
||||
|
||||
# 4) Antwortobjekte (Chunk-Ebene)
|
||||
results: List[QueryHit] = []
|
||||
for pid, total, e, c, s_score in scored[: req.top_k]:
|
||||
payload = id2payload[pid]
|
||||
results.append(QueryHit(
|
||||
node_id=pid,
|
||||
note_id=payload.get("note_id"),
|
||||
semantic_score=float(s_score),
|
||||
edge_bonus=float(e),
|
||||
centrality_bonus=float(c),
|
||||
total_score=float(total),
|
||||
paths=None,
|
||||
source={"path": payload.get("path"), "section": payload.get("section_title")}
|
||||
))
|
||||
|
||||
dt = int((time.time() - t0) * 1000)
|
||||
return QueryResponse(results=results, used_mode="hybrid", latency_ms=dt)
|
||||
# explain Flag durchreichen
|
||||
return _build_hits_from_semantic(
|
||||
hits,
|
||||
top_k=top_k,
|
||||
used_mode="hybrid",
|
||||
subgraph=subgraph,
|
||||
explain=req.explain
|
||||
)
|
||||
@@ -0,0 +1,116 @@
|
||||
"""app/core/retriever_config.py
|
||||
---------------------------------
|
||||
Zentrale Konfiguration für den mindnet-Retriever (WP-04).
|
||||
|
||||
Zweck:
|
||||
- Lädt config/retriever.yaml (falls vorhanden) oder nutzt sinnvolle Defaults.
|
||||
- Bietet einen gecachten Zugriff auf die Retriever-Config für
|
||||
andere Module (z. B. graph_adapter, retriever).
|
||||
|
||||
Hinweis zur Weiterentwicklung (Selbstjustierung):
|
||||
- Die hier definierten Parameter sind so gewählt, dass sie später
|
||||
durch ein Feedback-/Learning-to-Rank-Modell überschrieben werden
|
||||
können, ohne die restliche Architektur anzupassen.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import os
|
||||
from dataclasses import dataclass
|
||||
from functools import lru_cache
|
||||
from pathlib import Path
|
||||
from typing import Dict
|
||||
|
||||
try:
|
||||
import yaml # type: ignore
|
||||
except Exception: # pragma: no cover - Fallback, falls PyYAML nicht installiert ist.
|
||||
yaml = None # type: ignore
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class RetrieverConfig:
|
||||
semantic_scale: float
|
||||
edge_scale: float
|
||||
centrality_scale: float
|
||||
edge_weights: Dict[str, float]
|
||||
|
||||
@lru_cache
|
||||
def get_retriever_config() -> RetrieverConfig:
|
||||
"""Lädt die Retriever-Konfiguration (YAML + Defaults).
|
||||
|
||||
Reihenfolge:
|
||||
1. Defaults (sinnvoll gewählte Startwerte).
|
||||
2. Optional: config/retriever.yaml bzw. Pfad aus ENV
|
||||
MINDNET_RETRIEVER_CONFIG überschreibt die Defaults.
|
||||
|
||||
Die Funktion ist bewusst gecached, da sich die Konfiguration zur
|
||||
Laufzeit üblicherweise nicht ändert. Für dynamisches Nachladen
|
||||
könnte der Cache explizit geleert werden.
|
||||
"""
|
||||
|
||||
# 1) Defaults – bewusst konservativ gewählt.
|
||||
semantic_scale = 1.0
|
||||
edge_scale = 1.0
|
||||
centrality_scale = 1.0
|
||||
|
||||
edge_weights: Dict[str, float] = {
|
||||
# Wissens-Kanten
|
||||
"depends_on": 1.0,
|
||||
"related_to": 0.7,
|
||||
"similar_to": 0.7,
|
||||
"references": 0.5,
|
||||
# Struktur-Kanten
|
||||
"belongs_to": 0.2,
|
||||
"next": 0.1,
|
||||
"prev": 0.1,
|
||||
# Sonstige / technische Kanten
|
||||
"backlink": 0.2,
|
||||
"references_at": 0.2,
|
||||
}
|
||||
|
||||
# 2) Optional: YAML-Konfiguration laden
|
||||
cfg_path_env = os.getenv("MINDNET_RETRIEVER_CONFIG")
|
||||
if cfg_path_env:
|
||||
cfg_path = Path(cfg_path_env)
|
||||
else:
|
||||
# Project-Root = zwei Ebenen über app/core/
|
||||
cfg_path = Path(__file__).resolve().parents[2] / "config" / "retriever.yaml"
|
||||
|
||||
if yaml is not None and cfg_path.exists():
|
||||
try:
|
||||
data = yaml.safe_load(cfg_path.read_text(encoding="utf-8")) or {}
|
||||
except Exception:
|
||||
data = {}
|
||||
|
||||
retr = data.get("retriever") or {}
|
||||
|
||||
# Skalenwerte überschreiben, falls angegeben
|
||||
try:
|
||||
semantic_scale = float(retr.get("semantic_scale", semantic_scale))
|
||||
except (TypeError, ValueError):
|
||||
pass
|
||||
|
||||
try:
|
||||
edge_scale = float(retr.get("edge_scale", edge_scale))
|
||||
except (TypeError, ValueError):
|
||||
pass
|
||||
|
||||
try:
|
||||
centrality_scale = float(retr.get("centrality_scale", centrality_scale))
|
||||
except (TypeError, ValueError):
|
||||
pass
|
||||
|
||||
# Edge-Gewichte je Kanten-Typ
|
||||
ew_cfg = retr.get("edge_weights") or {}
|
||||
if isinstance(ew_cfg, dict):
|
||||
for k, v in ew_cfg.items():
|
||||
try:
|
||||
edge_weights[str(k)] = float(v)
|
||||
except (TypeError, ValueError):
|
||||
continue
|
||||
|
||||
return RetrieverConfig(
|
||||
semantic_scale=semantic_scale,
|
||||
edge_scale=edge_scale,
|
||||
centrality_scale=centrality_scale,
|
||||
edge_weights=edge_weights,
|
||||
)
|
||||
+63
-11
@@ -3,27 +3,27 @@ app/models/dto.py — Pydantic-Modelle (DTOs) für WP-04 Endpunkte
|
||||
|
||||
Zweck:
|
||||
Laufzeit-Modelle für FastAPI (Requests/Responses), getrennt von JSON-Schemas.
|
||||
Deckt die Graph-/Retriever-Endpunkte ab, ohne bestehende Funktionen zu ändern.
|
||||
Deckt die Graph-/Retriever-Endpunkte ab.
|
||||
Enthält Erweiterungen für WP-04b (Explanation Layer).
|
||||
|
||||
Kompatibilität:
|
||||
Python 3.12+, Pydantic 2.x, FastAPI 0.110+
|
||||
Version:
|
||||
0.1.0 (Erstanlage für WP-04)
|
||||
0.2.0 (Update für WP-04b Explanation Layer)
|
||||
Stand:
|
||||
2025-10-07
|
||||
2025-12-07
|
||||
Bezug:
|
||||
- schemas/*.json (Speicherschema für Notes/Chunks/Edges)
|
||||
- WP-04 API-Design (Query- und Graph-Endpunkte)
|
||||
Nutzung:
|
||||
from app.models.dto import QueryRequest, QueryResponse, GraphResponse
|
||||
Änderungsverlauf:
|
||||
0.1.0 (2025-10-07) – Erstanlage.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
from pydantic import BaseModel
|
||||
from typing import List, Literal, Optional, Dict
|
||||
from pydantic import BaseModel, Field
|
||||
from typing import List, Literal, Optional, Dict, Any
|
||||
|
||||
EdgeKind = Literal["references", "references_at", "backlink", "next", "prev", "belongs_to"]
|
||||
EdgeKind = Literal["references", "references_at", "backlink", "next", "prev", "belongs_to", "depends_on", "related_to", "similar_to"]
|
||||
|
||||
|
||||
class NodeDTO(BaseModel):
|
||||
@@ -44,7 +44,7 @@ class NodeDTO(BaseModel):
|
||||
class EdgeDTO(BaseModel):
|
||||
"""Darstellung einer Kante im API-Graph."""
|
||||
id: str
|
||||
kind: EdgeKind
|
||||
kind: str # String statt Literal, um flexibel für Custom-Types zu bleiben
|
||||
source: str
|
||||
target: str
|
||||
weight: float
|
||||
@@ -57,26 +57,78 @@ class QueryRequest(BaseModel):
|
||||
- mode: 'semantic' | 'edge' | 'hybrid'
|
||||
- query: (optional) Freitext; Embedding wird später angebunden
|
||||
- query_vector: (optional) direkter Vektor (384d) für Quick-Tests ohne Embedding
|
||||
- explain: (optional) Fordert detaillierte Erklärungen an (WP-04b)
|
||||
"""
|
||||
mode: Literal["semantic", "edge", "hybrid"] = "hybrid"
|
||||
query: Optional[str] = None
|
||||
query_vector: Optional[List[float]] = None
|
||||
top_k: int = 10
|
||||
expand: Dict = {"depth": 1, "edge_types": ["references", "belongs_to", "prev", "next"]}
|
||||
expand: Dict = {"depth": 1, "edge_types": ["references", "belongs_to", "prev", "next", "depends_on", "related_to"]}
|
||||
filters: Optional[Dict] = None
|
||||
# Flags zur Steuerung der Rückgabe
|
||||
ret: Dict = {"with_paths": True, "with_notes": True, "with_chunks": True}
|
||||
# WP-04b: Soll eine Erklärung generiert werden?
|
||||
explain: bool = False
|
||||
|
||||
|
||||
# --- WP-04b Explanation Models ---
|
||||
|
||||
class ScoreBreakdown(BaseModel):
|
||||
"""
|
||||
Aufschlüsselung der Score-Komponenten.
|
||||
Zeigt die gewichteten Beiträge zum Total Score.
|
||||
"""
|
||||
semantic_contribution: float = Field(..., description="W_sem * semantic_score * weight")
|
||||
edge_contribution: float = Field(..., description="W_edge * edge_bonus")
|
||||
centrality_contribution: float = Field(..., description="W_cent * centrality_bonus")
|
||||
|
||||
# Rohwerte für Transparenz
|
||||
raw_semantic: float
|
||||
raw_edge_bonus: float
|
||||
raw_centrality: float
|
||||
node_weight: float = Field(..., description="Typ-Gewicht (retriever_weight)")
|
||||
|
||||
|
||||
class Reason(BaseModel):
|
||||
"""
|
||||
Ein semantischer Grund für das Ranking.
|
||||
z.B. 'Verlinkt von Projekt X', 'Hohe Textähnlichkeit'.
|
||||
"""
|
||||
kind: Literal["semantic", "edge", "type", "centrality"]
|
||||
message: str
|
||||
score_impact: Optional[float] = None
|
||||
details: Optional[Dict[str, Any]] = None
|
||||
|
||||
|
||||
class Explanation(BaseModel):
|
||||
"""
|
||||
Container für alle Erklärungsdaten eines Treffers.
|
||||
"""
|
||||
breakdown: ScoreBreakdown
|
||||
reasons: List[Reason]
|
||||
# Optional: Pfade im Graphen, die zu diesem Treffer geführt haben
|
||||
related_edges: Optional[List[EdgeDTO]] = None
|
||||
|
||||
|
||||
# --- End Explanation Models ---
|
||||
|
||||
|
||||
class QueryHit(BaseModel):
|
||||
"""Einzelnes Trefferobjekt für /query."""
|
||||
node_id: str
|
||||
note_id: Optional[str]
|
||||
|
||||
# Flache Scores (Kompatibilität WP-04a)
|
||||
semantic_score: float
|
||||
edge_bonus: float
|
||||
centrality_bonus: float
|
||||
total_score: float
|
||||
|
||||
paths: Optional[List[List[Dict]]] = None
|
||||
source: Optional[Dict] = None
|
||||
|
||||
# WP-04b: Erklärungsobjekt (nur gefüllt, wenn explain=True)
|
||||
explanation: Optional[Explanation] = None
|
||||
|
||||
|
||||
class QueryResponse(BaseModel):
|
||||
@@ -91,4 +143,4 @@ class GraphResponse(BaseModel):
|
||||
center_note_id: str
|
||||
nodes: List[NodeDTO]
|
||||
edges: List[EdgeDTO]
|
||||
stats: Dict[str, int]
|
||||
stats: Dict[str, int]
|
||||
@@ -0,0 +1,24 @@
|
||||
version: 1.0
|
||||
|
||||
scoring:
|
||||
# W_sem: skaliert den Term (semantic_score * retriever_weight)
|
||||
# Empfehlung Startwert: 1.0 → Semantik bleibt Hauptsignal
|
||||
semantic_weight: 1.0
|
||||
|
||||
# W_edge: skaliert edge_bonus aus dem Subgraph
|
||||
# Empfehlung: 0.7 → Graph ist deutlich spürbar, aber überstimmt Semantik nicht komplett
|
||||
edge_weight: 0.7
|
||||
|
||||
# W_cent: skaliert centrality_bonus (Knoten-Zentralität im Subgraph)
|
||||
# Empfehlung: 0.5 → zentrale Knoten werden bevorzugt, aber moderat
|
||||
centrality_weight: 0.5
|
||||
|
||||
# Optional, für später (aktuell noch nicht ausgewertet, aber als Blaupause):
|
||||
edge_types:
|
||||
references: 0.20
|
||||
depends_on: 0.18
|
||||
related_to: 0.15
|
||||
similar_to: 0.12
|
||||
belongs_to: 0.10
|
||||
next: 0.06
|
||||
prev: 0.06
|
||||
+67
-9
@@ -1,18 +1,76 @@
|
||||
version: 1.0
|
||||
|
||||
defaults:
|
||||
retriever_weight: 1.0
|
||||
chunk_profile: default
|
||||
edge_defaults: [] # Leerliste → keine impliziten Kanten, wenn nicht gesetzt
|
||||
|
||||
types:
|
||||
# Wissensbausteine
|
||||
concept:
|
||||
chunk_profile: medium
|
||||
retriever_weight: 0.60
|
||||
edge_defaults: ["references", "related_to"]
|
||||
retriever_weight: 0.33
|
||||
task:
|
||||
|
||||
source:
|
||||
chunk_profile: short
|
||||
edge_defaults: ["depends_on", "belongs_to"]
|
||||
retriever_weight: 0.8
|
||||
experience:
|
||||
retriever_weight: 0.50
|
||||
edge_defaults: [] # Quellen sind i. d. R. Ziel von Kanten, erzeugen selbst keine Default-Kanten
|
||||
|
||||
glossary:
|
||||
chunk_profile: short
|
||||
retriever_weight: 0.40
|
||||
edge_defaults: ["related_to"]
|
||||
|
||||
# Person & Persönlichkeit
|
||||
profile: # z. B. "Persönliches Leitbild (Kurzprofil)"
|
||||
chunk_profile: long
|
||||
retriever_weight: 0.66
|
||||
edge_defaults: ["references", "related_to"]
|
||||
|
||||
value: # Werte (z. B. "Mut", "Fairness")
|
||||
chunk_profile: short
|
||||
retriever_weight: 0.90
|
||||
edge_defaults: ["related_to"]
|
||||
|
||||
principle: # Prinzipien/Leitsätze
|
||||
chunk_profile: short
|
||||
retriever_weight: 0.90
|
||||
edge_defaults: ["related_to", "references"]
|
||||
|
||||
experience: # prägende Erlebnisse/Erfahrungen
|
||||
chunk_profile: medium
|
||||
edge_defaults: ["derived_from", "inspired_by"]
|
||||
retriever_weight: 0.9
|
||||
retriever_weight: 0.90
|
||||
edge_defaults: ["related_to", "references"]
|
||||
|
||||
# Ziele, Entscheidungen, Meilensteine
|
||||
goal:
|
||||
chunk_profile: medium
|
||||
retriever_weight: 0.90
|
||||
edge_defaults: ["depends_on", "related_to", "references"]
|
||||
|
||||
decision:
|
||||
chunk_profile: short
|
||||
retriever_weight: 0.90
|
||||
edge_defaults: ["depends_on", "references", "related_to"] # „Worauf stützte sich die Entscheidung?“
|
||||
|
||||
milestone:
|
||||
chunk_profile: short
|
||||
retriever_weight: 0.70
|
||||
edge_defaults: ["related_to", "references"]
|
||||
|
||||
# Operativ: Projekte, Aufgaben, Journal
|
||||
project:
|
||||
chunk_profile: long
|
||||
edge_defaults: ["references", "depends_on"]
|
||||
retriever_weight: 0.95
|
||||
retriever_weight: 0.97
|
||||
edge_defaults: ["references", "depends_on"] # Projekt stützt sich auf Konzepte/Quellen und hängt von Bausteinen ab
|
||||
|
||||
task:
|
||||
chunk_profile: short
|
||||
retriever_weight: 0.80
|
||||
edge_defaults: ["depends_on", "references"]
|
||||
|
||||
journal:
|
||||
chunk_profile: medium
|
||||
retriever_weight: 0.80
|
||||
edge_defaults: ["references", "related_to"]
|
||||
@@ -0,0 +1,214 @@
|
||||
# mindnet v2.2 – Knowledge Design Manual
|
||||
**Datei:** `docs/mindnet_knowledge_design_manual_v2.2.md`
|
||||
**Stand:** 2025-12-07
|
||||
**Status:** **FINAL** (Konsolidiert aus WP01–WP04a)
|
||||
**Quellen:** `knowledge_design.md`, `TYPE_REGISTRY_MANUAL.md`, `chunking_strategy.md`, `mindnet_functional_architecture.md`, `WP03/WP04 Hinweise`.
|
||||
|
||||
---
|
||||
|
||||
## 1. Zweck & Scope
|
||||
|
||||
Dieses Handbuch ist die **primäre Arbeitsanweisung** für dich als Mindmaster (Owner) und alle Autoren, die Inhalte im mindnet-Vault erstellen.
|
||||
|
||||
### 1.1 Zielsetzung
|
||||
Mindnet ist mehr als eine Dokumentablage. Es ist ein vernetztes System, das deine Persönlichkeit, Entscheidungen und Erfahrungen abbildet. Damit der **Retriever** (die Suchmaschine) und spätere **KI-Agenten** dieses Wissen nutzen können, müssen Notizen einer klaren Struktur folgen.
|
||||
|
||||
### 1.2 Der Vault als „Source of Truth“
|
||||
Die Markdown-Dateien in deinem Vault sind die **einzige Quelle der Wahrheit**.
|
||||
* Mindnet importiert diesen Zustand deterministisch in die Datenbank (Qdrant).
|
||||
* Änderungen an IDs, Typen oder Inhalten müssen **immer** im Markdown erfolgen, niemals direkt in der Datenbank.
|
||||
* Was nicht im Markdown steht (z. B. implizites Wissen im Kopf), existiert für das System nicht.
|
||||
|
||||
---
|
||||
|
||||
## 2. Note-Struktur & Frontmatter
|
||||
|
||||
Jede Notiz benötigt einen YAML-Header (Frontmatter) am Dateianfang. Dieser Block macht die Notiz maschinenlesbar.
|
||||
|
||||
### 2.1 Pflichtfelder
|
||||
Jede Datei muss mindestens folgende Felder enthalten, um korrekt verarbeitet zu werden:
|
||||
|
||||
---
|
||||
id: 20251110-projekt-alpha-8a9b2c # Eindeutige Kennung
|
||||
title: Projekt Alpha # Sprechender Titel (wird in Suchergebnissen angezeigt)
|
||||
type: project # Steuert Verarbeitung & Vernetzung (siehe Kap. 3)
|
||||
status: active # Status (z.B. draft, active, archived)
|
||||
created: 2025-11-10 # Erstellungsdatum (ISO 8601 oder YYYY-MM-DD)
|
||||
updated: 2025-12-07 # Letzte Änderung
|
||||
tags: [ki, entwicklung] # Taxonomie für Filterung
|
||||
---
|
||||
|
||||
### 2.2 Optionale Felder
|
||||
Diese Felder sind technisch nicht zwingend, aber für bestimmte Typen sinnvoll:
|
||||
|
||||
lang: de # Sprache (Default: de)
|
||||
aliases: [Alpha Projekt, Project A] # Synonyme für die Suche
|
||||
visibility: internal # internal (default), public, private
|
||||
|
||||
> **Hinweis:** Felder wie `retriever_weight` oder `chunk_profile` sollten **nicht** mehr manuell im Frontmatter gesetzt werden. Diese werden zentral über den `type` gesteuert (siehe Kap. 3).
|
||||
|
||||
### 2.3 Empfehlungen für IDs und Pfade
|
||||
|
||||
**Die ID (Identifikator):**
|
||||
* Muss global eindeutig und **stabil** sein.
|
||||
* Darf sich nicht ändern, wenn die Datei umbenannt oder verschoben wird.
|
||||
* **Empfehlung:** `YYYYMMDD-slug-hash` (z. B. `20231027-vektor-db-a1b2`). Dies garantiert Eindeutigkeit und Chronologie.
|
||||
|
||||
**Dateinamen & Pfade:**
|
||||
* Pfade dienen der menschlichen Ordnung (Ordnerstruktur), sind für Mindnet aber sekundär.
|
||||
* Dateinamen sollten dem Titel entsprechen (`Projekt Alpha.md`) oder dem Slug (`projekt-alpha.md`).
|
||||
* Sonderzeichen vermeiden (außer Bindestrich und Unterstrich).
|
||||
|
||||
---
|
||||
|
||||
## 3. Note-Typen & Typ-Registry
|
||||
|
||||
Der `type` ist der wichtigste Hebel im Knowledge Design. Er verknüpft die Notiz mit der Konfiguration in `config/types.yaml`.
|
||||
|
||||
### 3.1 Übersicht der Kern-Typen
|
||||
|
||||
Mindnet unterscheidet verschiedene Wissensarten. Wähle den Typ, der die **Rolle** der Notiz am besten beschreibt:
|
||||
|
||||
| Typ | Beschreibung & Einsatzzweck |
|
||||
| :--- | :--- |
|
||||
| **`concept`** | Fachbegriffe, Theorien, Modelle. Zeitloses Wissen. (z. B. "Vektordatenbank") |
|
||||
| **`project`** | Ein Vorhaben mit Ziel, Dauer und Aufgaben. (z. B. "Aufbau Homelab") |
|
||||
| **`experience`** | Persönliche Erfahrung, Lektion oder Erkenntnis. (z. B. "Learning: Docker Networking") |
|
||||
| **`decision`** | Eine bewusst getroffene Entscheidung inkl. Alternativen und Begründung. (z. B. "ADR-001: Nutzung von Qdrant") |
|
||||
| **`value`** | Ein persönlicher Wert oder ein Prinzip. (z. B. "Datensparsamkeit") |
|
||||
| **`person`** | Eine reale Person (Netzwerk, Autor). |
|
||||
| **`event`** | Ein Ereignis (Konferenz, Meeting). |
|
||||
| **`journal`** | Zeitbezogener Log-Eintrag, Daily Note. |
|
||||
| **`source`** | Externe Quelle (Buch, PDF, Artikel), die exzerpiert wird. |
|
||||
|
||||
### 3.2 Zusammenspiel mit `types.yaml`
|
||||
|
||||
Der `type` steuert im Hintergrund drei technische Mechanismen:
|
||||
|
||||
1. **`retriever_weight` (Wichtigkeit):**
|
||||
* Ein `concept` (0.6) wiegt weniger als ein `project` (0.97) oder eine `decision` (1.0).
|
||||
* Ziel: Mindnet soll bei Fragen eher *deine Entscheidungen* und *Projekte* finden als reine Definitionen.
|
||||
2. **`chunk_profile` (Textzerlegung):**
|
||||
* `journal` (short): Wird fein zerlegt, da sich Themen schnell ändern.
|
||||
* `concept` (medium): Standard-Absätze.
|
||||
* `project` (long): Längere Kontext-Fenster, um Zusammenhänge zu wahren.
|
||||
3. **`edge_defaults` (Automatische Vernetzung):**
|
||||
* Mindnet ergänzt automatisch Kanten.
|
||||
* Beispiel: Ein Link in einem `project` wird automatisch als `depends_on` (Abhängigkeit) interpretiert, sofern nicht anders angegeben.
|
||||
|
||||
---
|
||||
|
||||
## 4. Edges & Referenzen in Notes
|
||||
|
||||
Um aus isolierten Dateien ein **Netzwerk** zu machen, nutzen wir Verlinkungen. Mindnet v2.2 (WP03) bietet hierfür präzise Werkzeuge.
|
||||
|
||||
### 4.1 Wikilinks (Die Basis-Referenz)
|
||||
Der klassische Obsidian-Link.
|
||||
|
||||
Wir nutzen [[Qdrant]] als Vektordatenbank.
|
||||
|
||||
* **Bedeutung:** "Diese Notiz erwähnt Qdrant."
|
||||
* **Edge-Typ:** `references`
|
||||
* **Confidence:** 1.0
|
||||
|
||||
### 4.2 Inline-Relationen (Semantische Verknüpfung)
|
||||
Wenn du ausdrücken willst, **wie** Dinge zusammenhängen. Dies ist essenziell für die Erklärbarkeit ("Warum ist das relevant?").
|
||||
|
||||
Daher [[rel:depends_on Qdrant]].
|
||||
Dieses Konzept ist [[rel:similar_to Pinecone]].
|
||||
|
||||
* **Syntax:** `[[rel:RELATION ZIEL]]`. (Wichtig: Das `rel:` Präfix steht *innerhalb* der Klammer).
|
||||
* **Gültige Relationen:**
|
||||
* `depends_on`: Hängt ab von / Benötigt.
|
||||
* `similar_to`: Ähnelt / Ist vergleichbar mit.
|
||||
* `related_to`: Hat zu tun mit (allgemein).
|
||||
* `caused_by`: Wurde verursacht durch.
|
||||
* `solves`: Löst (Problem).
|
||||
* `implements`: Setzt um (Konzept).
|
||||
* **Nicht unterstützt:** `rel: depends_on [[Ziel]]` (alte Syntax, funktioniert nicht mehr!).
|
||||
|
||||
### 4.3 Callout-Edges (Kuratierte Listen)
|
||||
Für Zusammenfassungen oder "Siehe auch"-Blöcke am Ende einer Notiz, die nicht im Fließtext stehen sollen.
|
||||
|
||||
> [!edge] related_to: [[Vector Embeddings]] [[AI Agents]]
|
||||
|
||||
* **Funktion:** Erzeugt `related_to`-Kanten zu allen genannten Zielen in dieser Zeile.
|
||||
* **Nutzen:** Erlaubt schnelle Listenpflege ohne den Textfluss zu stören.
|
||||
|
||||
### 4.4 Strukturkanten (Automatisch)
|
||||
Diese musst du nicht schreiben, sie entstehen automatisch:
|
||||
* `belongs_to`: Verknüpft jeden Textabschnitt (Chunk) mit seiner Ursprungsnotiz.
|
||||
* `next` / `prev`: Verknüpft Absätze in Lesereihenfolge (ermöglicht Agenten das "Weiterlesen").
|
||||
* `has_part` / `part_of`: (Zukünftig) Basierend auf Ordnerstrukturen.
|
||||
|
||||
---
|
||||
|
||||
## 5. Best Practices & Beispiele
|
||||
|
||||
### 5.1 Beispiel: Projekt-Notiz
|
||||
Projekte profitieren von `depends_on`, um Abhängigkeiten zu klären.
|
||||
|
||||
---
|
||||
id: 20251115-proj-mindnet
|
||||
title: Mindnet Implementierung
|
||||
type: project
|
||||
status: active
|
||||
---
|
||||
|
||||
# Mindnet Implementierung
|
||||
|
||||
Wir bauen ein persönliches Wissensnetz.
|
||||
|
||||
## Tech Stack
|
||||
Wir nutzen [[rel:depends_on Qdrant]] für die Vektorsuche und [[rel:depends_on FastAPI]] für das Backend.
|
||||
|
||||
## Architektur
|
||||
Das Konzept basiert auf [[RAG Architecture]]. (Automatisch 'depends_on' durch Typ-Default).
|
||||
|
||||
### 5.2 Beispiel: Entscheidung (Decision Record)
|
||||
Entscheidungen sind hoch gewichtet (`retriever_weight: 1.0`).
|
||||
|
||||
---
|
||||
id: 20251120-adr-vektordb
|
||||
title: ADR: Wahl von Qdrant
|
||||
type: decision
|
||||
status: final
|
||||
tags: [architektur, db]
|
||||
---
|
||||
|
||||
# Entscheidung: Qdrant
|
||||
|
||||
Wir haben uns für Qdrant entschieden.
|
||||
|
||||
## Alternativen
|
||||
Wir haben auch [[rel:similar_to Pinecone]] und [[rel:similar_to Weaviate]] betrachtet.
|
||||
|
||||
## Begründung
|
||||
Qdrant erlaubt lokalen Betrieb und [[rel:solves Payload Filtering Requirements]].
|
||||
|
||||
### 5.3 Typische Anti-Patterns (Bitte vermeiden!)
|
||||
|
||||
1. **Der "Alles"-Zettel:** Eine Notiz "Ideen 2025", die 50 verschiedene Themen enthält.
|
||||
* *Besser:* Atomare Notizen ("Idee A", "Idee B") erstellen und verlinken.
|
||||
2. **Manuelle Gewichtung:** `retriever_weight: 0.9` im Frontmatter setzen.
|
||||
* *Problem:* Wenn du das System tunen willst, musst du hunderte Dateien ändern.
|
||||
* *Besser:* Den `type` korrekt setzen und das Gewicht zentral in `types.yaml` steuern.
|
||||
3. **Link-Wüsten:** Eine Notiz, die nur aus einer Liste von 100 Links besteht ohne Text.
|
||||
* *Problem:* Der Vektor-Sucher findet keinen kontextuellen "Anker" (Text).
|
||||
* *Besser:* Kurze Beschreibung zu jedem Link hinzufügen.
|
||||
|
||||
---
|
||||
|
||||
## 6. Langfristige Stabilität & Virtual Schema Layer
|
||||
|
||||
Mindnet ist auf Langlebigkeit ausgelegt.
|
||||
|
||||
### 6.1 Das "Virtual Schema" Prinzip
|
||||
Wir vermeiden es, Logik in den Markdown-Dateien hart zu kodieren.
|
||||
* Statt in jeder Notiz zu schreiben `chunk_size: 500`, schreiben wir nur `type: essay`.
|
||||
* Wie groß ein Chunk für ein Essay ist, definieren wir in der Konfiguration (`types.yaml`).
|
||||
|
||||
### 6.2 Was bedeutet das für dich?
|
||||
* Du kannst dich auf den Inhalt konzentrieren.
|
||||
* Wenn wir in Zukunft entscheiden, dass "Projekte" stärker gewichtet werden sollen, ändern wir **eine Zeile** in der Konfiguration, und das gesamte System passt sich beim nächsten Import an.
|
||||
* Deine Notizen bleiben sauber und zukunftssicher.
|
||||
@@ -0,0 +1,190 @@
|
||||
# Mindnet v2.2 – Admin Guide
|
||||
**Datei:** `docs/mindnet_admin_guide_v2.2.md`
|
||||
**Stand:** 2025-12-07
|
||||
**Status:** **FINAL** (Konsolidiert WP02–WP04a)
|
||||
**Quellen:** `Handbuch.md`, `mindnet_v2_implementation_playbook.md`, `mindnet_technical_architecture.md`, `Programmplan_V2.2.md`.
|
||||
|
||||
> Dieses Handbuch richtet sich an **Administratoren**. Es beschreibt Installation, Konfiguration, Backup-Strategien, Monitoring und den sicheren Betrieb der Mindnet-Instanz.
|
||||
|
||||
---
|
||||
|
||||
## 1. Zielgruppe & Scope
|
||||
|
||||
Mindnet ist als **Single-Node System** konzipiert, das lokal (z.B. Laptop) oder auf einem privaten Server (Homelab, Intranet) läuft.
|
||||
Als Admin bist du verantwortlich für:
|
||||
* Die Verfügbarkeit der Dienste (Qdrant, API).
|
||||
* Die Datensicherheit (Backups).
|
||||
* Die Integrität des Imports (Sync zwischen Vault und DB).
|
||||
* Die Governance (wer darf zugreifen).
|
||||
|
||||
---
|
||||
|
||||
## 2. Initial Setup
|
||||
|
||||
### 2.1 Systemvoraussetzungen
|
||||
* **OS:** Linux (Ubuntu 22.04+ empfohlen) oder macOS.
|
||||
* **Runtime:** Python 3.10+, Docker (für Qdrant).
|
||||
* **Hardware:**
|
||||
* CPU: 2+ Cores.
|
||||
* RAM: Min. 4GB (abhängig von der Vault-Größe und Qdrant-Index).
|
||||
* Disk: SSD empfohlen für Qdrant-Performance.
|
||||
|
||||
### 2.2 Installation
|
||||
# 1. Repository klonen
|
||||
git clone <repo-url> /opt/mindnet
|
||||
cd /opt/mindnet
|
||||
|
||||
# 2. Umgebung einrichten
|
||||
python3 -m venv .venv
|
||||
source .venv/bin/activate
|
||||
pip install -r requirements.txt
|
||||
|
||||
# 3. Verzeichnisse anlegen
|
||||
mkdir -p logs qdrant_storage
|
||||
|
||||
### 2.3 Qdrant Setup (Docker)
|
||||
Wir nutzen Qdrant als Vektor-Datenbank. Persistenz ist wichtig.
|
||||
docker run -d \
|
||||
--name mindnet_qdrant \
|
||||
--restart always \
|
||||
-p 6333:6333 \
|
||||
-v $(pwd)/qdrant_storage:/qdrant/storage \
|
||||
qdrant/qdrant
|
||||
|
||||
### 2.4 Konfiguration (ENV)
|
||||
Erstelle eine `.env` Datei im Root-Verzeichnis. Diese Variablen steuern das Verhalten der Skripte und der API.
|
||||
|
||||
# Qdrant Verbindung
|
||||
QDRANT_URL="http://localhost:6333"
|
||||
QDRANT_API_KEY="" # Leer lassen für lokale Instanzen ohne Auth
|
||||
|
||||
# Mindnet Core Settings
|
||||
COLLECTION_PREFIX="mindnet"
|
||||
MINDNET_TYPES_FILE="./config/types.yaml"
|
||||
MINDNET_RETRIEVER_CONFIG="./config/retriever.yaml"
|
||||
|
||||
# Embedding Settings (Default: all-MiniLM-L6-v2)
|
||||
VECTOR_DIM=384
|
||||
|
||||
# Import-Strategie (Performance vs. Gründlichkeit)
|
||||
# Body: Nur Textänderungen triggern Update (Standard)
|
||||
# Full: Auch Frontmatter-Änderungen (z.B. Tags) triggern Update
|
||||
MINDNET_HASH_COMPARE="Body"
|
||||
MINDNET_HASH_SOURCE="parsed"
|
||||
|
||||
---
|
||||
|
||||
## 3. Betrieb im Alltag
|
||||
|
||||
### 3.1 Regelmäßige Importe
|
||||
Der Vault-Zustand sollte regelmäßig (z.B. stündlich per Cronjob oder Trigger) nach Qdrant synchronisiert werden. Das Skript erkennt Änderungen automatisch.
|
||||
|
||||
**Cronjob-Beispiel (stündlich):**
|
||||
0 * * * * cd /opt/mindnet && .venv/bin/python3 -m scripts.import_markdown --vault /path/to/vault --prefix "mindnet" --apply --purge-before-upsert --sync-deletes >> ./logs/import.log 2>&1
|
||||
|
||||
* `--purge-before-upsert`: Wichtig, um Fragmente gelöschter Textstellen zu entfernen.
|
||||
* `--sync-deletes`: Wichtig, um Notizen aus dem Index zu entfernen, die im Vault gelöscht wurden.
|
||||
|
||||
### 3.2 Health-Checks
|
||||
Prüfe regelmäßig, ob die API und der Retriever korrekt arbeiten.
|
||||
|
||||
**API Status Check:**
|
||||
curl -f http://localhost:8000/health || echo "API Down"
|
||||
|
||||
**Logischer Smoke-Test:**
|
||||
Führt eine echte Hybrid-Suche durch und prüft, ob Scores berechnet werden.
|
||||
python3 scripts/test_retriever_smoke.py --mode hybrid
|
||||
|
||||
### 3.3 Umgang mit Fehlern im Importer
|
||||
Falls der Importer abbricht oder Dateien ignoriert:
|
||||
|
||||
1. **Diagnose:** Nutze `scripts/diagnose_changed.py`, um zu sehen, warum eine Datei als "unchanged" gilt.
|
||||
python3 -m scripts.diagnose_changed --vault ./vault --note-id "MeineProblemnote"
|
||||
2. **Schema-Validierung:** Prüfe mit `scripts/payload_dryrun.py`, ob eine spezifische Datei das JSON-Schema verletzt.
|
||||
3. **Erzwingen:** Nutze `touch <filename>` im Vault, um den Zeitstempel zu aktualisieren und Hash-Checks zu umgehen, oder nutze `--force` (Vorsicht: Performance).
|
||||
|
||||
---
|
||||
|
||||
## 4. Update-Prozess
|
||||
|
||||
Wenn neue Versionen der Mindnet-Software (z.B. Schema-Updates in WP05) ausgerollt werden:
|
||||
|
||||
1. **Code aktualisieren:**
|
||||
git pull origin main
|
||||
2. **Dependencies prüfen:**
|
||||
pip install -r requirements.txt
|
||||
3. **Schema-Migration:**
|
||||
Mindnet nutzt oft "Rebuild" statt komplexer Migrationen, da der Vault die Source of Truth ist.
|
||||
* Prüfe `CHANGELOG.md` auf "Breaking Changes".
|
||||
* Bei Schema-Änderungen: Führe einen **Full Rebuild** durch (siehe 5.3).
|
||||
|
||||
---
|
||||
|
||||
## 5. Backup & Restore
|
||||
|
||||
Datensicherheit ruht auf zwei Säulen: Dem Vault (Original) und Qdrant (Index).
|
||||
|
||||
### 5.1 Vault-Backup (Priorität 1)
|
||||
Der Markdown-Vault ist die **Single Source of Truth**. Er muss klassisch gesichert werden.
|
||||
* **Strategie:** Git-Repository, rsync auf NAS, oder Cloud-Sync.
|
||||
* **Restore:** Einfach Dateien zurückkopieren und Import laufen lassen.
|
||||
|
||||
### 5.2 Qdrant-Snapshots (Priorität 2)
|
||||
Für schnelle Wiederherstellung des Suchindex ohne stundenlangen Re-Import (bei großen Vaults).
|
||||
|
||||
**Backup erstellen (Volume-Level):**
|
||||
docker stop mindnet_qdrant
|
||||
tar -czf qdrant_backup_$(date +%F).tar.gz ./qdrant_storage
|
||||
docker start mindnet_qdrant
|
||||
|
||||
**Alternativ: API-Snapshot (Zero-Downtime):**
|
||||
Siehe Qdrant Dokumentation (Snapshot API).
|
||||
|
||||
**Restore:**
|
||||
1. Container stoppen.
|
||||
2. Inhalt von `qdrant_storage` löschen.
|
||||
3. Backup entpacken.
|
||||
4. Container starten.
|
||||
|
||||
### 5.3 Notfall-Wiederherstellung (Rebuild)
|
||||
Wenn die Datenbank korrupt ist oder Indizes defekt sind, ist der sauberste Weg ein Rebuild:
|
||||
|
||||
# 1. DB komplett leeren (Wipe)
|
||||
python3 -m scripts.reset_qdrant --mode wipe --prefix "mindnet" --yes
|
||||
|
||||
# 2. Alles neu importieren
|
||||
python3 -m scripts.import_markdown --vault /path/to/vault --prefix "mindnet" --apply
|
||||
|
||||
---
|
||||
|
||||
## 6. Monitoring & Logging
|
||||
|
||||
### 6.1 Log-Quellen
|
||||
* **Importer:** Schreibt nach STDOUT/STDERR. Sollte im Cronjob in eine Datei (z.B. `/var/log/mindnet_import.log`) umgeleitet werden.
|
||||
* **API (Uvicorn):** Standard Access-Logs.
|
||||
* **Qdrant:** Docker Logs (`docker logs mindnet_qdrant`).
|
||||
|
||||
### 6.2 Wichtige Metriken (Log-Parsing)
|
||||
Achte in den Import-Logs auf folgende Muster:
|
||||
* `edges_created`: Sollte bei neuen Inhalten > 0 sein.
|
||||
* `chunks_created`: Sollte plausibel zur Anzahl neuer/geänderter Notizen sein.
|
||||
* `unresolved_references`: Ein hoher Anstieg deutet auf tote Links oder Tippfehler im Vault hin.
|
||||
* `ERROR` / `WARNING`: Parsing-Fehler (z.B. ungültiges YAML).
|
||||
|
||||
---
|
||||
|
||||
## 7. Governance & Sicherheit
|
||||
|
||||
### 7.1 Zugriffsschutz
|
||||
Mindnet hat aktuell **keine integrierte Authentifizierung** (By Design für lokalen Betrieb).
|
||||
* **API:** Muss hinter einem Reverse Proxy (Nginx, Traefik, Caddy) mit Basic Auth oder OAuth laufen, wenn sie im Netzwerk freigegeben wird.
|
||||
* **Qdrant:** Sollte nicht öffentlich erreichbar sein (Firewall auf Port 6333, `127.0.0.1` Bindung).
|
||||
|
||||
### 7.2 Datenschutz & Sichtbarkeit
|
||||
* **Interne Nutzung:** Standardmäßig sind alle Inhalte im Index "internal".
|
||||
* **Trennung:** Wenn du öffentliche und private Daten hast, nutze unterschiedliche `COLLECTION_PREFIX` (z.B. `mindnet_public`, `mindnet_private`) und getrennte Import-Jobs/Vault-Ordner, oder warte auf die WP05-ACL-Implementierung (Filterung auf API-Ebene).
|
||||
|
||||
### 7.3 Typen-Governance
|
||||
Änderungen an der `types.yaml` (z.B. neue Gewichte) wirken global.
|
||||
* **Prozess:** Änderungen sollten getestet werden (Smoke-Test), bevor sie produktiv gehen.
|
||||
* **Review:** Prüfe, ob neue Typen (`type: xyz`) auch Chunking-Profile haben, sonst fallen sie auf Defaults zurück.
|
||||
@@ -0,0 +1,140 @@
|
||||
# Mindnet v2.2 – Appendices & Referenzen
|
||||
**Datei:** `docs/mindnet_appendices_v2.2.md`
|
||||
**Stand:** 2025-12-07
|
||||
**Status:** **FINAL** (Konsolidiert aus WP01–WP04a)
|
||||
**Quellen:** `TYPE_REGISTRY_MANUAL.md`, `chunking_strategy.md`, `mindnet_technical_architecture.md`, `Handbuch.md`.
|
||||
|
||||
> Dieses Dokument bündelt Tabellen, Schemata und technische Referenzen, die in den Prozess-Dokumenten (Playbook, Guides) den Lesefluss stören würden.
|
||||
|
||||
---
|
||||
|
||||
## Anhang A: Typ-Registry Referenz (Default-Werte)
|
||||
|
||||
Diese Tabelle zeigt die Standard-Konfiguration der `types.yaml` (Stand v2.2). Diese Werte werden beim Import angewendet, sofern im Frontmatter nichts anderes definiert ist.
|
||||
|
||||
| Typ (`type`) | Chunk Profile | Retriever Weight | Edge Defaults (Auto-Kanten) | Beschreibung |
|
||||
| :--- | :--- | :--- | :--- | :--- |
|
||||
| **concept** | `medium` | 0.60 | `references`, `related_to` | Abstrakte Begriffe, Theorien. |
|
||||
| **project** | `long` | 0.97 | `references`, `depends_on` | Aktive Vorhaben. Hohe Priorität. |
|
||||
| **decision** | `long` | 1.00 | `caused_by`, `references` | Entscheidungen (ADRs). Höchste Prio. |
|
||||
| **experience** | `medium` | 0.90 | `derived_from`, `inspired_by` | Persönliche Learnings. |
|
||||
| **journal** | `short` | 0.80 | `related_to` | Zeitgebundene Logs. Fein granular. |
|
||||
| **person** | `short` | 0.50 | `related_to` | Personen-Profile. |
|
||||
| **source** | `long` | 0.50 | *(keine)* | Externe Quellen (Bücher, PDFs). |
|
||||
| **event** | `short` | 0.60 | `related_to` | Meetings, Konferenzen. |
|
||||
| **default** | `medium` | 1.00 | `references` | Fallback, wenn Typ unbekannt. |
|
||||
|
||||
---
|
||||
|
||||
## Anhang B: Edge-Typen & Semantik
|
||||
|
||||
Referenz aller implementierten Kantenarten (`kind`). Die Spalte "Herkunft" zeigt, durch welchen Mechanismus diese Kante primär entsteht.
|
||||
|
||||
| Kind | Herkunft (Primär) | Symmetrisch? | Bedeutung & Nutzung |
|
||||
| :--- | :--- | :--- | :--- |
|
||||
| `references` | Wikilink `[[...]]` | Nein | Standard-Verweis. "Erwähnt X". |
|
||||
| `belongs_to` | Struktur (Auto) | Nein | Verknüpft Chunk mit seiner Note. |
|
||||
| `next` / `prev` | Struktur (Auto) | Ja (Logisch) | Definiert Lesereihenfolge der Chunks. |
|
||||
| `depends_on` | Inline / Default | Nein | Harte Abhängigkeit. "Braucht X um zu funktionieren". |
|
||||
| `related_to` | Callout / Default | Ja (Oft) | Thematische Nähe. "Siehe auch X". |
|
||||
| `similar_to` | Inline | Ja | Inhaltliche Ähnlichkeit. "Ist wie X". |
|
||||
| `caused_by` | Inline | Nein | Kausalität. "X ist der Grund für Y". |
|
||||
| `solves` | Inline | Nein | Lösung. "Tool X löst Problem Y". |
|
||||
| `derived_from` | Default (Exp) | Nein | Herkunft. "Erkenntnis stammt aus Quelle X". |
|
||||
|
||||
---
|
||||
|
||||
## Anhang C: Datenmodelle (JSON Payloads)
|
||||
|
||||
Dies sind die Felder, die effektiv in Qdrant gespeichert werden. Dient als Referenz für Entwickler und Debugging (`payload_dryrun`).
|
||||
|
||||
### C.1 Note Payload (`mindnet_notes`)
|
||||
{
|
||||
"note_id": "string (keyword)", // UUIDv5 oder Slug
|
||||
"title": "string (text)", // Titel aus Frontmatter
|
||||
"type": "string (keyword)", // Typ (z.B. 'project')
|
||||
"retriever_weight": "float", // Numerische Wichtigkeit (0.0-1.0)
|
||||
"chunk_profile": "string", // Genutztes Profil (z.B. 'long')
|
||||
"edge_defaults": ["string"], // Liste der aktiven Default-Kanten
|
||||
"tags": ["string"], // Liste von Tags
|
||||
"created": "string (iso-date)", // Erstellungsdatum
|
||||
"updated": "integer", // Timestamp
|
||||
"fulltext": "string (no-index)" // Gesamter Text (für Recovery)
|
||||
}
|
||||
|
||||
### C.2 Chunk Payload (`mindnet_chunks`)
|
||||
{
|
||||
"chunk_id": "string (keyword)", // Format: {note_id}#c{index}
|
||||
"note_id": "string (keyword)", // FK zur Note
|
||||
"text": "string (text)", // Reintext für Anzeige (ohne Overlap)
|
||||
"window": "string (text)", // Text + Overlap (für Embedding)
|
||||
"ord": "integer", // Laufende Nummer (1..N)
|
||||
"retriever_weight": "float", // Kopie aus Note (für Query-Speed)
|
||||
"neighbors_prev": ["string"], // ID des Vorgängers
|
||||
"neighbors_next": ["string"] // ID des Nachfolgers
|
||||
}
|
||||
|
||||
### C.3 Edge Payload (`mindnet_edges`)
|
||||
{
|
||||
"edge_id": "string (keyword)", // Deterministischer Hash
|
||||
"source_id": "string (keyword)", // Chunk-ID (Start)
|
||||
"target_id": "string (keyword)", // Chunk-ID oder Note-Titel (Ziel)
|
||||
"kind": "string (keyword)", // z.B. 'depends_on'
|
||||
"scope": "string (keyword)", // Immer 'chunk' in v2.2
|
||||
"rule_id": "string (keyword)", // Traceability: 'inline:rel', 'explicit:wikilink'
|
||||
"confidence": "float", // 0.0 - 1.0
|
||||
"note_id": "string (keyword)" // Owner Note ID (für Löschung)
|
||||
}
|
||||
|
||||
---
|
||||
|
||||
## Anhang D: Environment Variablen
|
||||
|
||||
Diese Variablen steuern das Verhalten der Skripte und Container.
|
||||
|
||||
| Variable | Default | Beschreibung |
|
||||
| :--- | :--- | :--- |
|
||||
| `QDRANT_URL` | `http://localhost:6333` | URL zur Vektor-DB. |
|
||||
| `QDRANT_API_KEY` | *(leer)* | API-Key für Absicherung (optional). |
|
||||
| `COLLECTION_PREFIX` | `mindnet` | Namensraum für Collections (`{prefix}_notes` etc). |
|
||||
| `MINDNET_TYPES_FILE` | `config/types.yaml` | Pfad zur Typ-Registry. |
|
||||
| `MINDNET_RETRIEVER_CONFIG`| `config/retriever.yaml`| Pfad zur Scoring-Konfiguration. |
|
||||
| `MINDNET_HASH_COMPARE` | `Body` | Vergleichsmodus für Import (`Body`, `Frontmatter`, `Full`). |
|
||||
| `MINDNET_HASH_SOURCE` | `parsed` | Quelle für Hash (`parsed`, `raw`, `file`). |
|
||||
| `MINDNET_HASH_NORMALIZE` | `canonical` | Normalisierung vor Hash (`canonical`, `none`). |
|
||||
| `VECTOR_DIM` | `384` | Dimension der Embeddings (Modellabhängig). |
|
||||
|
||||
---
|
||||
|
||||
## Anhang E: Glossar
|
||||
|
||||
Zentrale Begriffe des Mindnet-Systems.
|
||||
|
||||
* **Callout-Edge:** Eine Kante, die durch einen Obsidian-Callout Block (`> [!edge]`) definiert wird. Dient für Listen.
|
||||
* **Chunking:** Prozess des Zerlegens einer Notiz in kleinere, semantische Einheiten (Chunks).
|
||||
* **Confidence:** Ein Wert (0.0-1.0), der angibt, wie sicher eine Kante ist. Explizite User-Eingaben haben 1.0, Defaults ~0.7.
|
||||
* **Idempotenz:** Eigenschaft des Importers, bei mehrfacher Ausführung das exakt gleiche Ergebnis zu liefern (keine Duplikate).
|
||||
* **Inline-Edge:** Eine Kante, die im Fließtext mittels `[[rel:type Ziel]]` definiert wird. Semantisch sehr stark.
|
||||
* **Provenance:** Die Herkunft einer Information (z.B. "explicit" vs "rule" vs "inference").
|
||||
* **Retriever:** Die Komponente (FastAPI), die Suchanfragen beantwortet und Scores berechnet.
|
||||
* **Rule ID:** Ein String, der genau identifiziert, welche Code-Regel eine Kante erzeugt hat (z.B. `edge_defaults:project:depends_on`).
|
||||
* **Vault:** Der lokale Ordner mit Markdown-Dateien. Die "Single Source of Truth".
|
||||
* **Vector Search:** Suche basierend auf semantischer Ähnlichkeit (Embeddings) statt reiner Stichworte.
|
||||
|
||||
---
|
||||
|
||||
## Anhang F: Workpackage Status (v2.2)
|
||||
|
||||
Status der Module gemäß Programmplan.
|
||||
|
||||
| WP | Titel | Status | Anmerkung |
|
||||
| :--- | :--- | :--- | :--- |
|
||||
| **WP01** | Knowledge Design | 🟢 Live | Typen, Frontmatter definiert. |
|
||||
| **WP02** | Chunking Strategy | 🟢 Live | Profilbasiertes Chunking aktiv. |
|
||||
| **WP03** | Edge Logic / Import | 🟢 Live | Neue Importer-Pipeline mit Provenance. |
|
||||
| **WP04a**| Retriever Scoring | 🟢 Live | Hybrider Score (Semantik + Graph). |
|
||||
| **WP04b**| Explanation Layer | 🟡 In Arbeit | API liefert Pfade, UI fehlt noch. |
|
||||
| **WP04c**| Feedback Loop | 🔴 Geplant | Logging von User-Feedback. |
|
||||
| **WP05** | Virtual Schema | 🟡 Teilw. | `types.yaml` aktiv, ACL fehlt. |
|
||||
| **WP06** | Self-Healing | 🔴 Geplant | Auto-Fixing von Broken Links. |
|
||||
| **WP08** | Self-Tuning | 🔴 Geplant | Auto-Adjustment der Gewichte. |
|
||||
@@ -0,0 +1,143 @@
|
||||
# Mindnet v2.2 – Entwickler-Workflow
|
||||
**Datei:** `DEV_WORKFLOW.md`
|
||||
**Stand:** 2025-12-07
|
||||
|
||||
Dieses Handbuch beschreibt den Entwicklungszyklus zwischen **Windows PC** (IDE), **Raspberry Pi** (Gitea) und **Beelink** (Runtime/Server).
|
||||
|
||||
---
|
||||
|
||||
## 1. Die Architektur
|
||||
|
||||
* **Windows 11 (VS Code):** Hier schreibst du Code. **Nie** direkt auf `main` arbeiten!
|
||||
* **Raspberry Pi (Gitea):** Der Tresor. Speichert den Code und verwaltet Versionen.
|
||||
* **Beelink (Runtime):** Hier läuft die Software. Wir trennen strikt:
|
||||
* **PROD (`~/mindnet`):** Läuft stabil auf `main`. **Port 8001**.
|
||||
* **DEV (`~/mindnet_dev`):** Deine Spielwiese für Branches. **Port 8002**.
|
||||
|
||||
---
|
||||
|
||||
## 2. Der Zyklus: Von der Idee zum Release
|
||||
|
||||
### Phase 1: Entwicklung (Windows / VS Code)
|
||||
|
||||
Hier erstellst du die neue Funktion in einer sicheren Umgebung.
|
||||
|
||||
1. **Sicherstellen, dass Git bereit ist:**
|
||||
* Öffne VS Code.
|
||||
* Unten links in der blauen Leiste sollte der aktuelle Branch stehen (z.B. `main`).
|
||||
|
||||
2. **Branch erstellen (WICHTIG):**
|
||||
* Klicke unten links auf den Branch-Namen (z.B. `main`).
|
||||
* Wähle im Menü oben `+ Create new branch...`.
|
||||
* Gib den Namen ein: `feature/was-ich-tue` (z.B. `feature/wp04b-explanation`).
|
||||
* Drücke **Enter**.
|
||||
|
||||
3. **Sicherheits-Check:**
|
||||
* Schau unten links: Steht dort jetzt `feature/wp04b-explanation`?
|
||||
* **Nur wenn dort dein Feature-Branch steht, darfst du Code ändern!** Alles, was du jetzt speicherst, landet in diesem Branch.
|
||||
|
||||
4. **Coden:**
|
||||
* Nimm deine Änderungen vor.
|
||||
|
||||
5. **Sichern & Hochladen:**
|
||||
* Klicke links auf das **Source Control** Icon (Gabel-Symbol).
|
||||
* Gib eine Nachricht ein: "feat: Add explanation layer logic".
|
||||
* Klicke auf **Commit**.
|
||||
* Klicke auf **Publish Branch** (oder "Sync Changes"), um den Branch zum Raspberry Pi (Gitea) zu senden.
|
||||
|
||||
---
|
||||
|
||||
### Phase 2: Testen (Beelink / `mindnet_dev`)
|
||||
|
||||
Hier prüfst du, ob dein neuer Code auf dem echten Server läuft, ohne die Produktion zu stören.
|
||||
|
||||
1. **Einloggen:**
|
||||
`ssh user@<BEELINK-IP>`
|
||||
|
||||
2. **In die Werkstatt gehen:**
|
||||
`cd /home/llmadmin/mindnet_dev`
|
||||
|
||||
3. **Code holen:**
|
||||
```bash
|
||||
git fetch
|
||||
git checkout feature/wp04b-explanation
|
||||
git pull
|
||||
```
|
||||
|
||||
4. **Umgebung aktivieren (Venv):**
|
||||
Damit Python die installierten Bibliotheken findet, musst du das Virtual Environment aktivieren:
|
||||
```bash
|
||||
source .venv/bin/activate
|
||||
```
|
||||
*(Dein Prompt sollte jetzt `(.venv) user@...` zeigen)*
|
||||
|
||||
5. **Abhängigkeiten prüfen:**
|
||||
*(Nur nötig, wenn du neue Libraries in `requirements.txt` eingetragen hast)*
|
||||
```bash
|
||||
pip install -r requirements.txt
|
||||
```
|
||||
|
||||
6. **Test-Server starten:**
|
||||
**WICHTIG:** Nutze Port **8002** und das Dev-Prefix!
|
||||
Stelle sicher, dass in `~/mindnet_dev/.env` steht: `COLLECTION_PREFIX="mindnet_dev"`.
|
||||
|
||||
```bash
|
||||
uvicorn app.main:app --host 0.0.0.0 --port 8002
|
||||
```
|
||||
|
||||
7. **Testen:**
|
||||
Prüfe deine Änderung (z.B. Browser auf `http://<BEELINK-IP>:8002/docs` oder Smoke-Test Skript).
|
||||
|
||||
---
|
||||
|
||||
### Phase 3: Release (Gitea / Raspberry Pi)
|
||||
|
||||
Wenn der Test auf Port 8002 erfolgreich war:
|
||||
|
||||
1. Öffne Gitea im Browser.
|
||||
2. Du solltest einen Hinweis sehen "feature/... had recent pushes".
|
||||
3. Erstelle einen **Pull Request** (Dein Branch -> `main`).
|
||||
4. **Merge** den Pull Request.
|
||||
|
||||
---
|
||||
|
||||
### Phase 4: Deployment (Beelink / `mindnet`)
|
||||
|
||||
Jetzt bringen wir die Änderung in das Live-System (Port 8001).
|
||||
|
||||
1. **Deploy-Agent (Automatisch):**
|
||||
Wenn dein `deploy.yml` aktiv ist, passiert das jetzt von selbst!
|
||||
|
||||
2. **Manuell (Falls nötig):**
|
||||
```bash
|
||||
cd /home/llmadmin/mindnet
|
||||
git pull origin main
|
||||
sudo systemctl restart mindnet-api
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 3. Cheat Sheet: Die wichtigsten Befehle
|
||||
|
||||
| Wo? | Befehl | Was tut es? |
|
||||
| :--- | :--- | :--- |
|
||||
| **VS Code** | `Klick auf Branch-Namen` | Branch erstellen oder wechseln. |
|
||||
| **VS Code** | `Sync Changes` | Lädt Code zu Gitea hoch. |
|
||||
| **Beelink** | `cd ~/mindnet_dev` | Gehe in den Test-Ordner. |
|
||||
| **Beelink** | `source .venv/bin/activate` | **Aktiviert Python-Umgebung.** |
|
||||
| **Beelink** | `git fetch` | "Gibt es neue Branches auf Gitea?" |
|
||||
| **Beelink** | `git checkout <name>` | Wechsle auf einen anderen Branch. |
|
||||
| **Beelink** | `git pull` | Hol den neuesten Stand des aktuellen Branches. |
|
||||
| **Beelink** | `uvicorn ... --port 8002` | Startet Test-Server (Dev). |
|
||||
|
||||
---
|
||||
|
||||
## 4. Sicherheitsregeln
|
||||
|
||||
1. **Niemals** in `~/mindnet` (Prod-Ordner) experimentieren oder `git checkout` machen. Dieser Ordner bleibt immer auf `main`.
|
||||
2. **Niemals** beide Umgebungen auf denselben `COLLECTION_PREFIX` zeigen lassen.
|
||||
* Prod `.env`: `COLLECTION_PREFIX="mindnet"`
|
||||
* Dev `.env`: `COLLECTION_PREFIX="mindnet_dev"`
|
||||
3. **Port-Disziplin:**
|
||||
* **PROD:** Port 8001
|
||||
* **DEV:** Port 8002
|
||||
@@ -0,0 +1,155 @@
|
||||
# Mindnet v2.2 – Developer Guide
|
||||
**Datei:** `docs/mindnet_developer_guide_v2.2.md`
|
||||
**Stand:** 2025-12-07
|
||||
**Status:** **FINAL** (Konsolidiert WP02–WP04a)
|
||||
**Quellen:** `mindnet_technical_architecture.md`, `Handbuch.md`, `docs_mindnet_retriever.md`, `mindnet_v2_implementation_playbook.md`.
|
||||
|
||||
> **Zielgruppe:** Entwickler:innen.
|
||||
> **Zweck:** Anleitung zum Aufsetzen der Entwicklungsumgebung, Verständnis der Modulstruktur und Durchführung von Tests.
|
||||
|
||||
---
|
||||
|
||||
## 1. Projektstruktur
|
||||
|
||||
Der Code ist modular in `app` (Logik), `scripts` (CLI) und `config` (Steuerung) getrennt.
|
||||
|
||||
mindnet/
|
||||
├── app/
|
||||
│ ├── core/ # Kernlogik
|
||||
│ │ ├── chunker.py # Text-Zerlegung
|
||||
│ │ ├── derive_edges.py # Edge-Erzeugung (WP03 Logik)
|
||||
│ │ ├── retriever.py # Scoring & Graph-Expansion (WP04 Logik)
|
||||
│ │ ├── qdrant.py # DB-Verbindung
|
||||
│ │ └── ...
|
||||
│ ├── models/ # Pydantic DTOs (Request/Response)
|
||||
│ ├── routers/ # FastAPI Endpoints
|
||||
│ ├── services/ # Externe Dienste (LLM, Embeddings)
|
||||
│ └── main.py # Entrypoint der API
|
||||
├── config/ # YAML-Konfigurationen (Single Source of Truth)
|
||||
│ ├── types.yaml # Import-Regeln
|
||||
│ └── retriever.yaml # Scoring-Regeln
|
||||
├── scripts/ # CLI-Tools (Import, Diagnose, Reset)
|
||||
├── tests/ # Pytest Suite
|
||||
└── vault/ # Dein lokaler Markdown-Content (Git-ignored)
|
||||
|
||||
---
|
||||
|
||||
## 2. Lokales Setup (Development)
|
||||
|
||||
### 2.1 Voraussetzungen
|
||||
* **Python:** 3.10 oder höher.
|
||||
* **Docker:** Für Qdrant.
|
||||
* **Vault:** Ein Ordner mit Markdown-Dateien (z.B. `./mindnet_v2_test_vault` für Tests).
|
||||
|
||||
### 2.2 Installation
|
||||
# 1. Repository klonen & Verzeichnis wechseln
|
||||
git clone <repo> mindnet
|
||||
cd mindnet
|
||||
|
||||
# 2. Virtual Environment erstellen
|
||||
python3 -m venv .venv
|
||||
source .venv/bin/activate
|
||||
|
||||
# 3. Abhängigkeiten installieren
|
||||
pip install -r requirements.txt
|
||||
|
||||
### 2.3 Konfiguration (Environment)
|
||||
Erstelle eine `.env` Datei im Root-Verzeichnis oder exportiere die Variablen.
|
||||
|
||||
# Qdrant Verbindung
|
||||
QDRANT_URL="http://localhost:6333"
|
||||
QDRANT_API_KEY="" # Leer lassen für lokal
|
||||
|
||||
# Mindnet Settings
|
||||
COLLECTION_PREFIX="mindnet_dev"
|
||||
MINDNET_TYPES_FILE="./config/types.yaml"
|
||||
MINDNET_RETRIEVER_CONFIG="./config/retriever.yaml"
|
||||
|
||||
# Hash-Strategie (für Importer-Debugging)
|
||||
MINDNET_HASH_COMPARE="Body"
|
||||
MINDNET_HASH_SOURCE="parsed"
|
||||
|
||||
### 2.4 Dienste starten
|
||||
Starten von Qdrant via Docker:
|
||||
docker run -d --name mindnet_qdrant_dev -p 6333:6333 -p 6334:6334 qdrant/qdrant
|
||||
|
||||
---
|
||||
|
||||
## 3. Core-Module & Entwicklung
|
||||
|
||||
### 3.1 Der Importer (`scripts.import_markdown`)
|
||||
Dies ist das komplexeste Modul.
|
||||
* **Einstieg:** `scripts/import_markdown.py` -> `main()`.
|
||||
* **Idempotenz:** Der Importer muss mehrfach laufen können, ohne Duplikate zu erzeugen. Wir nutzen deterministische IDs (UUIDv5) basierend auf Dateipfaden/Titeln.
|
||||
* **Debugging:** Wenn du am Importer arbeitest, nutze `--dry-run` oder `scripts/payload_dryrun.py`, um DB-Schreibvorgänge zu vermeiden.
|
||||
|
||||
### 3.2 Edge-Logik (`app.core.derive_edges`)
|
||||
Hier wird entschieden, welche Kanten entstehen.
|
||||
* **Erweiterung:** Wenn du neue Edge-Regeln (z.B. Regex-Parser) hinzufügst, tue dies hier.
|
||||
* **Rule-ID:** Vergib zwingend eine eindeutige `rule_id` (z.B. `custom:my_rule`), damit die Herkunft nachvollziehbar bleibt.
|
||||
|
||||
### 3.3 Der Retriever (`app.core.retriever`)
|
||||
Hier passiert das Scoring und die Graph-Expansion.
|
||||
* **Gewichtung:** Hardcodiere niemals Gewichte (`0.5` etc.) im Code. Nutze `app.core.config.get_retriever_config()`, um `retriever.yaml` zu lesen.
|
||||
* **Graph-Adapter:** Nutze `graph_adapter.expand()`, um Nachbarn zu laden. Vermeide direkte Qdrant-Point-Queries im Retriever-Code, um die Abstraktion zu wahren.
|
||||
|
||||
---
|
||||
|
||||
## 4. Tests & Debugging
|
||||
|
||||
Wir unterscheiden drei Test-Ebenen. Ein Pull Request sollte alle passieren.
|
||||
|
||||
### 4.1 Unit Tests (Pytest)
|
||||
Für isolierte Logik (Parsing, Scoring).
|
||||
pytest tests/test_retriever_basic.py
|
||||
pytest tests/test_chunking.py
|
||||
pytest tests/test_edges_all.py
|
||||
|
||||
### 4.2 Integration / Pipeline Tests
|
||||
Prüfen den Datenfluss von Markdown bis Qdrant-JSON.
|
||||
* **Payload Dryrun:** Prüft, ob `import_markdown` valide JSON-Objekte erzeugt, ohne in die DB zu schreiben. Essentiell bei Schema-Änderungen.
|
||||
python3 -m scripts.payload_dryrun --vault ./mindnet_v2_test_vault
|
||||
|
||||
* **Edge Checks:** Prüft Graph-Invarianten (z.B. `next` muss reziprok zu `prev` sein).
|
||||
python3 -m scripts.edges_full_check
|
||||
|
||||
### 4.3 Smoke Tests (E2E)
|
||||
Prüfen das laufende System (API) gegen eine echte (lokale) Qdrant-Instanz.
|
||||
# 1. API im Hintergrund starten
|
||||
uvicorn app.main:app --port 8000 &
|
||||
|
||||
# 2. Smoke Test feuern
|
||||
python scripts/test_retriever_smoke.py --query "Test" --mode hybrid --top-k 5
|
||||
|
||||
---
|
||||
|
||||
## 5. Guidelines für Erweiterungen
|
||||
|
||||
### 5.1 Neuen Note-Typ hinzufügen
|
||||
* **Nicht im Code!** Bearbeite `config/types.yaml`.
|
||||
* Definiere `chunk_profile` und `retriever_weight`.
|
||||
* Führe einen **Re-Import** durch, damit die Änderungen wirksam werden.
|
||||
|
||||
### 5.2 Schema-Änderungen
|
||||
Wenn du Felder in `mindnet_notes` oder `mindnet_chunks` hinzufügst:
|
||||
1. Passe `app/core/note_payload.py` bzw. `chunk_payload.py` an.
|
||||
2. Aktualisiere `tests/assert_payload_schema.py`.
|
||||
3. Führe `scripts/reset_qdrant.py` (Wipe) aus, um die Indizes neu anzulegen.
|
||||
|
||||
### 5.3 Neue API-Endpunkte
|
||||
* Erstelle einen neuen Router in `app/routers/`.
|
||||
* Nutze Pydantic Models aus `app/models/dto.py` für Request/Response.
|
||||
* Registriere den Router in `app/main.py`.
|
||||
|
||||
---
|
||||
|
||||
## 6. Nützliche Einzeiler
|
||||
|
||||
**DB komplett zurücksetzen (Vorsicht!):**
|
||||
python3 -m scripts.reset_qdrant --mode wipe --prefix "mindnet_dev" --yes
|
||||
|
||||
**Einen einzelnen File inspizieren (Parser-Sicht):**
|
||||
python3 tests/inspect_one_note.py --file ./vault/MeinFile.md
|
||||
|
||||
**Warum wird mein File nicht importiert?**
|
||||
python3 -m scripts.diagnose_changed --vault ./vault --all
|
||||
@@ -0,0 +1,96 @@
|
||||
# mindnet Retriever (WP-04 / Step 4a) – Kurz-Dokumentation
|
||||
|
||||
## 1. Überblick
|
||||
|
||||
Der Retriever besteht aus:
|
||||
|
||||
- **Semantic Retriever** (`semantic_retrieve`)
|
||||
- **Hybrid Retriever** (`hybrid_retrieve`)
|
||||
- Qdrant-Backend (`*_chunks`, `*_edges`)
|
||||
- optionaler Edge-Expansion über `graph_adapter.expand`
|
||||
|
||||
Die API nutzt das bestehende `/query`-Endpoint (FastAPI), das `QueryRequest` und `QueryResponse` aus `app.models.dto` verwendet.
|
||||
|
||||
## 2. Scoring-Formel
|
||||
|
||||
Pro Treffer werden folgende Komponenten geführt:
|
||||
|
||||
- `semantic_score` – direkt aus Qdrant (Cosine/Euclidean, je nach Collection)
|
||||
- `retriever_weight` – aus Chunk-Payload (`types.yaml` / Frontmatter)
|
||||
- `edge_bonus` – Bonus aus dem Subgraph (z. B. Zahl der eingehenden/ausgehenden Edges, Beziehungstypen)
|
||||
- `centrality_bonus` – Bonus aus einer einfachen Zentralitäts-Heuristik im Subgraph
|
||||
|
||||
Aktuelle Formel (Step 4a):
|
||||
|
||||
```text
|
||||
total_score = semantic_score * max(retriever_weight, 0.0)
|
||||
+ edge_bonus
|
||||
+ centrality_bonus
|
||||
```
|
||||
|
||||
Die Sortierung der Ergebnisse erfolgt absteigend nach `total_score`.
|
||||
|
||||
## 3. Modi
|
||||
|
||||
### 3.1 Semantic Mode
|
||||
|
||||
- Funktion: `semantic_retrieve`
|
||||
- Request-Feld `mode = "semantic"`
|
||||
- Keine Edge-Expansion, kein Graph-Call.
|
||||
- Kandidaten nur aus Vektorsuche in `<prefix>_chunks`.
|
||||
|
||||
### 3.2 Hybrid Mode
|
||||
|
||||
- Funktion: `hybrid_retrieve`
|
||||
- Request-Feld `mode = "hybrid"`
|
||||
- Schritte:
|
||||
1. Semantische Chunk-Suche wie im Semantic Mode
|
||||
2. Falls `expand.depth > 0` gesetzt ist:
|
||||
- Seeds = `chunk_id` (Fallback: `note_id`) der Treffer
|
||||
- Aufruf `graph_adapter.expand(client, prefix, seed_ids, depth, edge_types)`
|
||||
- Ergebnis-Subgraph liefert `edge_bonus` und `centrality_bonus` pro Node
|
||||
3. Scores werden nach obiger Formel kombiniert.
|
||||
|
||||
## 4. Request-Struktur (JSON)
|
||||
|
||||
Beispiel-Hybrid-Request:
|
||||
|
||||
```json
|
||||
{
|
||||
"mode": "hybrid",
|
||||
"query": "embeddings und qdrant",
|
||||
"top_k": 10,
|
||||
"expand": {
|
||||
"depth": 1,
|
||||
"edge_types": ["references", "depends_on"]
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
Minimaler Semantic-Request:
|
||||
|
||||
```json
|
||||
{
|
||||
"mode": "semantic",
|
||||
"query": "karate trainingsplan",
|
||||
"top_k": 5
|
||||
}
|
||||
```
|
||||
|
||||
## 5. Smoke-Test
|
||||
|
||||
Für einen schnellen End-to-End-Test kann folgendes Skript verwendet werden:
|
||||
|
||||
```bash
|
||||
python scripts/test_retriever_smoke.py --query "karate" --mode hybrid --expand-depth 1 --top-k 5
|
||||
```
|
||||
|
||||
Das Skript ruft `/query` auf, wertet Statuscode und JSON-Antwort aus und zeigt die Scores der Top-K-Treffer an.
|
||||
|
||||
## 6. Tests
|
||||
|
||||
Die wichtigsten Tests für den Retriever sind:
|
||||
|
||||
- `tests/test_retriever_basic.py` – Basisfunktion (Semantic + Hybrid)
|
||||
- `tests/test_retriever_weight.py` – Einfluss von `retriever_weight` auf das Ranking
|
||||
- `tests/test_retriever_edges.py` – Einfluss von Edge-/Centrality-Boni im Hybrid-Modus
|
||||
+155
-188
@@ -1,249 +1,216 @@
|
||||
# mindnet – Knowledge Design
|
||||
**Version:** 1.4.0 (aktualisiert 2025-10-06)
|
||||
**Version:** 1.5.0 (aktualisiert 2025-11-09)
|
||||
|
||||
## 1. Gesamtziel und Systemverständnis
|
||||
|
||||
**mindnet** ist ein lokales Wissensnetzwerk, das alle meine Notizen, Projekte, Gedanken, Erfahrungen und Pläne als Markdown-Dateien speichert, automatisch importiert, semantisch in **Chunks** zerlegt und als **Graph-Struktur in Qdrant** ablegt.
|
||||
Es ist die Grundlage für ein agentenfähiges, lokal betriebenes LLM-System, das meine Daten verknüpft, abfragt und erweitert.
|
||||
|
||||
**Zentrale Ziele**
|
||||
- Vollständige Repräsentation persönlichen Wissens in strukturierter, maschinenlesbarer Form
|
||||
- Reproduzierbare Verarbeitung: Import → Chunking → Embedding → Edges → Graph
|
||||
- Verlustfreie Roundtrip-Fähigkeit (Import ⇄ Export)
|
||||
- Erweiterbarkeit für Agenten, RAG-Retriever und semantische Suche
|
||||
|
||||
**Prinzipien**
|
||||
- **Deterministisch:** IDs, Hashes und Edges wiederholbar
|
||||
- **Portabel:** Markdown + YAML
|
||||
- **Robust:** fehlertoleranter Parser
|
||||
- **Idempotent:** keine Duplikate bei Re-Import
|
||||
- **Erweiterbar:** neue Typen und Edge-Arten ohne Migration
|
||||
> Dieses Dokument beschreibt das Ziel, die Architektur, die Datenmodelle sowie die Regeln und Prozesse für den Aufbau von **mindnet** – einem persönlichen, lokal betriebenen Wissensnetz, das auf Markdown-Quellen, Graph-Beziehungen und Vektor-Suche basiert. Es soll die Persönlichkeit, Werte und Erfahrungen des Besitzers widerspiegeln und künftig Antworten generieren, die diese Perspektive authentisch berücksichtigen.
|
||||
|
||||
---
|
||||
|
||||
## 2. YAML-Frontmatter-Schema
|
||||
## 1. Ziel von mindnet
|
||||
|
||||
| Feld | Typ | Beschreibung |
|
||||
|-------|-----|--------------|
|
||||
| `title` | string | Menschlich lesbarer Titel |
|
||||
| `id` | string | Eindeutige Note-ID (Slug) |
|
||||
| `type` | enum | `concept`, `thought`, `experience`, `task`, `project`, `journal`, `person`, `meeting`, `milestone` |
|
||||
| `status` | enum | `draft`, `active`, `done`, `archived` |
|
||||
| `created` | date | ISO-Datum |
|
||||
| `updated` | date | Letzte Änderung |
|
||||
| `area` | string | Themen- oder Lebensbereich |
|
||||
| `project` | string | Zugehöriges Projekt |
|
||||
| `tags` | list | Strukturierte Tags (`area/…`, `topic/…`) |
|
||||
| `depends_on` | list | IDs anderer Notizen |
|
||||
| `assigned_to` | list | Beteiligte Personen |
|
||||
| `embedding_exclude` | bool | Falls `true`, keine Embeddings |
|
||||
| `hash_mode` | enum | Hash-Modus beim Import (`body`, `frontmatter`, `full`) |
|
||||
| `priority` | enum/int | 1–5 oder `low` / `med` / `high` |
|
||||
| `effort_min`, `due` | int/date | Aufwand / Termin |
|
||||
| `aliases` | list | Alternative IDs |
|
||||
| `lang` | string | ISO-Sprachcode |
|
||||
| `source` | string | Herkunft / Referenzquelle |
|
||||
**1.1 Zweck**
|
||||
- mindnet dient als persönliches Wissensnetz, das Erfahrungen, Werte, Prinzipien, Ziele und prägenden Ereignisse langfristig strukturiert.
|
||||
- Es ermöglicht spätere Einsichten in Entscheidungen und Entwicklungen und kann in Zukunft für die Familie (z. B. Kinder) eine nachvollziehbare Quelle sein.
|
||||
|
||||
**Pflichtfelder:** `title`, `id`, `type`, `status`, `created`
|
||||
**Optionale Felder:** alle übrigen
|
||||
**1.2 Leitidee**
|
||||
- Statt einer reinen Dokumentablage wird Wissen durch **Knoten (Notes/Chunks)** und **Kanten (Edges)** vernetzt.
|
||||
- Antworten sollen **persönlich** kontextualisiert werden (Perspektive, Konditionierung, Persönlichkeit).
|
||||
|
||||
**1.3 Langfristigkeit**
|
||||
- Stabiler, zukunftsfähiger Entwurf; spätere Massenmigrationen sollen vermieden werden.
|
||||
- Architektur unterstützt schrittweise Erweiterung (neue Quellen, neue Agenten, neue Regeln).
|
||||
|
||||
---
|
||||
|
||||
## 3. Dateinamen und Ordner
|
||||
## 2. Zielarchitektur (High Level)
|
||||
|
||||
- Format: `YYYY-MM-DD_title.md` oder `slug.md`
|
||||
- Strukturierte Ordner:
|
||||
- `10_thoughts/`, `20_concepts/`, `30_projects/`, `40_experiences/`, …
|
||||
- Alle Pfade relativ; Unterstriche statt Leerzeichen
|
||||
- **Datenquellen:** Markdown-Vault (Obsidian-kompatibel), perspektivisch weitere Quellen (MediaWiki, PDFs, Web-Exporte, E-Mail-Auszüge).
|
||||
- **Import-Pipeline:** Parser → Frontmatter/Body → Chunking → Embedding → Qdrant (Notes/Chunks/Edges).
|
||||
- **Speicher:** Qdrant (Vektoren + Payload + Textindex), Dateien bleiben im Filesystem (Quelle ist Referenz).
|
||||
- **Graph-Schicht:** Edges explizit modelliert (Typ, Richtung, Evidenz, Confidence); optionale abgeleitete Edges per Batch-Job.
|
||||
- **Retriever/Composer:** Hybrid-Retrieval (Vektor + Filter + optional Volltext), Komposition zu Antwortkandidaten.
|
||||
- **Erklärbarkeit:** Begründungspfade (Top-Chunks, Graph-Kette, Provenienz).
|
||||
- **Schnittstellen:** REST/CLI, n8n-Glue, lokale LLM-Anbindung (Ollama).
|
||||
|
||||
---
|
||||
|
||||
## 4. Link- und Edge-Design
|
||||
## 3. Collections in Qdrant
|
||||
|
||||
**Verlinkungen**
|
||||
- Wikilinks: `[[note-id]]` oder `[[Titel|note-id]]`
|
||||
- Automatische Edges aus Text und Frontmatter
|
||||
|
||||
**Edge-Typen in Qdrant**
|
||||
|
||||
| Typ | Bedeutung | Quelle | Scope |
|
||||
|------|------------|--------|--------|
|
||||
| `belongs_to` | Chunk → Note | intern | chunk |
|
||||
| `prev` / `next` | Reihenfolge | intern | chunk |
|
||||
| `references` | explizite Wikilinks | Text | chunk |
|
||||
| `backlink` | Gegenkante zu `references` | abgeleitet | note |
|
||||
| `depends_on` | YAML-Abhängigkeiten | YAML | note |
|
||||
| `assigned_to` | YAML-Zuweisung | YAML | note |
|
||||
| `unresolved` | Ziel fehlt (Stub) | abgeleitet | chunk |
|
||||
|
||||
**Edge-Schlüssel**
|
||||
- `source_id`, `target_id`, `kind`, `scope`, `note_id`
|
||||
- Dedup-Schlüssel: Kombination aus `(kind, source_id, target_id, scope)`
|
||||
- `mindnet_notes` – Note-Metadaten (id, title, type, path, timestamps, tags, …).
|
||||
- `mindnet_chunks` – inhaltstragende Segmente inkl. Vektoren und Textindex.
|
||||
- `mindnet_edges` – Beziehungen (src/dst, relation, direction, evidence, confidence, …).
|
||||
- (optional) `mindnet_people` – strukturierte Personenobjekte (Eigen-/Fremdbezug).
|
||||
|
||||
---
|
||||
|
||||
## 5. Chunking-Modell
|
||||
## 4. Identitäten & IDs
|
||||
|
||||
**Chunk-Datenstruktur**
|
||||
| Feld | Beschreibung |
|
||||
|-------|---------------|
|
||||
| `note_id` | Zugehörige Note |
|
||||
| `chunk_index` | Laufende Nummer |
|
||||
| `text` | Originalabschnitt |
|
||||
| `window` | Text + Overlap links/rechts |
|
||||
| `overlap_left` | Länge des linken Overlaps |
|
||||
| `overlap_right` | Länge des rechten Overlaps |
|
||||
| `embedding` | 384-dimensionaler Vektor |
|
||||
| `tokens` | Tokenanzahl (optional) |
|
||||
|
||||
**Chunking-Regeln (aus `chunking_strategy.md`)**
|
||||
- Trennung nach Absätzen, Überschriften, Listen, Tabellen
|
||||
- Token-Zielgröße 350–500 (max 600)
|
||||
- Overlap: 30–40 % für semantische Kohärenz
|
||||
- Overlap wird in `window` integriert, sodass `text ≠ window`
|
||||
- **Deterministische IDs** für Notes (z. B. Namespace-basierte UUIDv5) und stabile, ableitbare IDs für Chunks/Edges.
|
||||
- ID-Stabilität ist Voraussetzung für Idempotenz (Re-Imports ohne Duplikate).
|
||||
|
||||
---
|
||||
|
||||
## 6. Hash-Strategie und Änderungsverfolgung
|
||||
## 5. Frontmatter & Typen
|
||||
|
||||
**Zweck:** Nur bei realer Inhaltsänderung re-indexieren
|
||||
|
||||
| Umgebungsvariable | Beschreibung |
|
||||
|--------------------|--------------|
|
||||
| `MINDNET_HASH_COMPARE` | Vergleichsquelle (`Body`, `Full`, `Frontmatter`) |
|
||||
| `MINDNET_HASH_SOURCE` | Rohtext vs. geparst (`raw` / `parsed`) |
|
||||
| `MINDNET_HASH_NORMALIZE` | Normalisierung (`canonical` / `none`) |
|
||||
|
||||
**Basismodi**
|
||||
- `body` → Nur Textkörper
|
||||
- `frontmatter` → Nur Metadaten
|
||||
- `full` → Kombination
|
||||
- Hash = SHA-256(canonicalized_input)
|
||||
|
||||
Bei Änderung: Re-Import, neue Embeddings, Edges werden aktualisiert.
|
||||
- Pflichtfelder: `id`, `title`, `type`, `created`, `updated`, `tags`, `source_path`.
|
||||
- Note-Typen (Auszug): `experience`, `principle`, `value`, `goal`, `role`, `persona`, `plan`, `project`, `journal`, `reference`, `translation`.
|
||||
- **types.yaml** definiert Regeln/Defaults je Typ und für Relationstypen (siehe Abschnitt 10).
|
||||
|
||||
---
|
||||
|
||||
## 7. Qdrant-Speicherstruktur
|
||||
## 6. Chunking-Strategie
|
||||
|
||||
| Collection | Inhalt | Primärfelder | Embedding |
|
||||
|-------------|---------|---------------|------------|
|
||||
| `mindnet_notes` | Notes mit Fulltext & Hash | `note_id` | optional |
|
||||
| `mindnet_chunks` | Chunks mit Text, Window, Embedding | `chunk_id`, `note_id` | 384 d |
|
||||
| `mindnet_edges` | Graphkanten | `edge_id`, `source_id`, `target_id` | — |
|
||||
|
||||
**Upsert-Strategie**
|
||||
- Idempotente UUIDv5-IDs
|
||||
- Keine Duplikate durch deterministische Signaturen
|
||||
- `purge-before-upsert` zur Bereinigung
|
||||
|
||||
**Indizes**
|
||||
- notes: `note_id`, `hash_signature`
|
||||
- chunks: `note_id`, `chunk_index`
|
||||
- edges: `(kind, source_id, target_id, scope)`
|
||||
- Satz- bis Absatz-Chunks, moderate Überlappung; Meta-Felder verweisen auf Note/Quelle.
|
||||
- Named Vectors (z. B. `content`, optional `title`, `tags`) ermöglichen hybride Scores.
|
||||
|
||||
---
|
||||
|
||||
## 8. Fehler-Toleranz und Parser-Robustheit
|
||||
## 7. Embeddings & Vektoren
|
||||
|
||||
- UTF-8 mit Fallback: ersetzt ungültige Bytes (`errors="replace"`)
|
||||
- Ignoriert Nullbytes, BOM, exotische Zeilenenden
|
||||
- Erkennt defekte oder leere YAML-Header und überspringt sie
|
||||
- Logt Problemfälle (`read_markdown failed`, `make_note_payload returned non-dict`)
|
||||
- Bei Fehlern: Import → warn → continue
|
||||
- Einheitliche Metrik (cosine), reproduzierbares Embedding-Setup.
|
||||
- Konsistenz über Zeit durch Versionsfeld (`embedding_model`, `embedding_version`).
|
||||
|
||||
---
|
||||
|
||||
## 9. Tests & Validierung
|
||||
## 8. Volltext & Filter
|
||||
|
||||
**Integrationstests**
|
||||
- Roundtrip: Import → Export → Compare (`compare_vaults.py`)
|
||||
- Chunk-Integrität: `verify_chunks_integrity.py`
|
||||
- Hash-Audit: `hash_reporter.py`
|
||||
- Window-Vergleich: `check_chunks_window_vs_text.py`
|
||||
|
||||
**Erwartete Ergebnisse**
|
||||
- Alle Notes werden erkannt
|
||||
- `window ≠ text` bei > 60 % der Chunks
|
||||
- Roundtrip-Vergleich liefert „OK“
|
||||
- Edge-Anzahl entspricht Modell (belongs_to = #Chunks, next/prev ≈ #Chunks – 1)
|
||||
- Textindex auf `mindnet_chunks.body` für exakte Term/Operator-Queries.
|
||||
- Payload-Filter zur schnellen Einschränkung (tags, types, time-buckets, visibility u. a.).
|
||||
|
||||
---
|
||||
|
||||
## 10. Export & Roundtrip-Verhalten
|
||||
## 9. Edges – Beziehungen
|
||||
|
||||
- `export_markdown.py` erzeugt identische Markdown-Dateien
|
||||
- Schreibt YAML + Body in korrekter Reihenfolge
|
||||
- Roundtrip-Prüfung (`compare_vaults.py`) validiert:
|
||||
- vollständigen Body
|
||||
- unveränderte Hash-Signatur
|
||||
- gleiche Chunkzahl
|
||||
- **Explizit**: Jede tatsächliche Beziehung wird als Kante gespeichert (kein Workaround).
|
||||
- Felder u. a.: `src_id`, `dst_id`, `relation`, `direction`, `confidence (0–1)`, `evidence` (Chunk-IDs), `created_at`, `updated_at`.
|
||||
- **edge_defaults** liefern **Interpretations-Hints** (Gewichte, Symmetrie, Transitivität) und parametrisieren abgeleitete Edges; sie **ersetzen** keine Originalkanten.
|
||||
|
||||
---
|
||||
|
||||
## 11. Graph-Schicht und WP-04-Integration
|
||||
## 10. Regeln & Policies (types.yaml)
|
||||
|
||||
- Qdrant fungiert als Graph-Backend
|
||||
- `graph/service.py` (WP-04): stellt API für Mehrhop-Abfragen bereit
|
||||
- Unterstützt:
|
||||
- `expand(note_id, hops=2)`
|
||||
- `resolve_unresolved()`
|
||||
- `neighbors(kind=…)`
|
||||
- Grundlage für spätere **Retriever-Funktionen** (LLM / Agenten)
|
||||
- Per Typ/Relation:
|
||||
- Default-Eigenschaften (Gewicht, Directedness, Symmetrie, Transitivität).
|
||||
- Scoring-Hooks für Graph-Kontext im Retriever.
|
||||
- (später) Privacy-Hooks je Relation (Mindest-Visibility).
|
||||
|
||||
---
|
||||
|
||||
## 12. Agenten- und LLM-Integration
|
||||
## 11. Retrieval & Ranking (Komposition)
|
||||
|
||||
**Retriever-Pipeline**
|
||||
1. Suche top-k Chunks (cosine distance)
|
||||
2. Erweiterung um Edges (`references`, `backlinks`)
|
||||
3. Kontext-Assembling via `window`-Text
|
||||
4. Prompt-Generierung
|
||||
|
||||
**Agent-Use-Cases**
|
||||
- Automatisches Tagging (`type`, `status`)
|
||||
- Vorschläge für neue Edges
|
||||
- Auflösen von `unresolved`-Referenzen
|
||||
- Auto-Import externer Quellen (MediaWiki, PDFs, Webseiten)
|
||||
- Hybrid: Vektor-Score + Filter + optional BM25/Volltext.
|
||||
- Zeitgewicht (Decays) und Graph-Kontext werden komponiert (Formel siehe 16.3).
|
||||
- Konfiguration über `.env`/`config.yaml` (gewichts- und halbwertszeit-basiert).
|
||||
|
||||
---
|
||||
|
||||
## 13. Erweiterbarkeit
|
||||
## 12. Erklärbarkeit
|
||||
|
||||
- Neue Edge-Typen: `inspired_by`, `relates_to`, `contradicts`
|
||||
- Neue Note-Typen: `decision`, `resource`, `recommendation`
|
||||
- Embedding-Engines austauschbar (MiniLM → LaBSE → E5)
|
||||
- Versionierte Chunking-Strategien (`chunk_config.py`)
|
||||
- „Warum-Pfad“: Top-Chunks mit Score-Zerlegung (Vektor/Time/Graph), Graph-Kette, Provenienz.
|
||||
- Mehrschichtige Narrative (Kindheit/Erlebnisse, Werte/Prinzipien, Ziele/Leitbild, Gegenwart).
|
||||
|
||||
---
|
||||
|
||||
## 14. Qualitätssicherung und Best Practices
|
||||
## 13. Betrieb & Sicherheit
|
||||
|
||||
- Validierung vor Upsert
|
||||
- Keine Leer-Chunks
|
||||
- Vollständige Metadaten in Notes
|
||||
- Tests in `tests/` automatisch ausführbar (`run_e2e_roundtrip.sh`)
|
||||
- Ergebnis: „Roundtrip OK“ und `verify_chunks_integrity` = OK
|
||||
- Lokal auf Single-Node starten (Backups via Snapshots).
|
||||
- Zugriff nur intern/Reverse-Proxy; Festplattenverschlüsselung OS-seitig.
|
||||
- Migrationspfad zu Qdrant-Distributed (Shards, Replikation) einplanen.
|
||||
|
||||
---
|
||||
|
||||
## 15. Änderungsverlauf
|
||||
## 14. Qualität & Monitoring
|
||||
|
||||
**v1.4.0 (2025-10-06)**
|
||||
- Neu: `window` vs `text` Feld in `mindnet_chunks`
|
||||
- Neu: Hash-Modi (`body`, `frontmatter`, `full`) mit Env-Steuerung
|
||||
- Neu: Baseline-Modus für Vergleich
|
||||
- Parser → fehlertolerant (UTF-8 replace)
|
||||
- Roundtrip-Test vollständig integriert
|
||||
- Graph-Service und Mehrhop-Abfragen (WP-04-Vorbereitung)
|
||||
- `purge-before-upsert` und `--note-scope-refs` Parameter ergänzt
|
||||
|
||||
**v1.3.0 (2025-09-09)**
|
||||
- Hash-Normalisierung und Edge-Scope
|
||||
- Roundtrip-Tests, Agenten-Integration
|
||||
|
||||
**v1.2.0 (2025-09-02)**
|
||||
- Grundstruktur des Knowledge-Designs
|
||||
- Offline-Eval-Sets, Telemetrie (lokal), Drift-Checks, manuelle Reviews.
|
||||
- Kennzahlen: Hit-Rate, Overlap, Quellen-Alter, Erklärungspfad-Nutzung.
|
||||
|
||||
---
|
||||
|
||||
## 15. Offene Punkte
|
||||
|
||||
- Ausdetaillierung `types.yaml` (Relationen, Hooks).
|
||||
- UI-Oberfläche für Erklärpfade und Provenienz.
|
||||
- Batch-Jobs für abgeleitete Edges.
|
||||
|
||||
---
|
||||
|
||||
## 16. Ergänzungen 2025-11-09 (v1.5.0)
|
||||
|
||||
**Kontext:** Basierend auf den neuen Rahmenbedingungen (nur Deutsch; perspektivische externe Abfragen; einheitliche Vertraulichkeit mit Option auf feinere Stufen; Recency-Priorisierung als Default; Regeln später konfigurierbar; zunächst lokale Speicherung mit optionaler Migration; heutige Hardware, später skalierbar; hohe Erklärbarkeit mit mehreren Erklärungsebenen) werden folgende Entscheidungen und Erweiterungen getroffen. Alle bisherigen Inhalte bleiben gültig und werden **nicht** verändert.
|
||||
|
||||
### 16.1 Sprach- und Inhaltsrichtlinie (nur Deutsch)
|
||||
- `lang: "de"` als verpflichtendes Feld in allen Note-, Chunk- und Edge-Payloads.
|
||||
- Für spätere Mehrsprachigkeit **nicht** die Collection splitten, sondern ein Feld `lang` + optional `lang_original` behalten; Translation-Artefakte als eigener `type: translation` (verweisend auf `source_id`).
|
||||
|
||||
### 16.2 Vertraulichkeit & Sichtbarkeit
|
||||
- Heute: einheitliche Stufe `visibility: "internal"`.
|
||||
- Vorausschauend: reservierte Felder in allen Payloads
|
||||
- `visibility`: `"public" | "internal" | "restricted"`
|
||||
- `sensitivity`: Integer 0–3 (0 = unkritisch, 3 = hochsensibel)
|
||||
- `acl`: Liste erlaubter Rollen/Personen-IDs (App-Layer erzwingt Filter vor Query).
|
||||
- Access-Control erfolgt **im Applikations-Layer** mittels Qdrant-Payload-Filtern (`must: [ { key: "visibility", match: { any: [...] }}, ... ]`); Qdrant bleibt „dumm“ in Bezug auf ACL.
|
||||
|
||||
### 16.3 Recency-Modell & Ranking (konfigurierbar)
|
||||
- Default: **Zeitnähe** priorisieren (sanfte Abwertung älterer Inhalte).
|
||||
- Im Retrieval kombinieren wir:
|
||||
1) Vektor-Score (cosine) aus Qdrant
|
||||
2) Zeitgewicht `time_decay = exp( - ln(2) * age_days / half_life_days )`
|
||||
3) Graph-Kontext aus Edges (z. B. Degree, Path-Hops zum Anker „Ich/Persona“)
|
||||
- Konfigurierbare Gewichte (per `.env` oder `config.yaml`):
|
||||
```text
|
||||
FINAL_SCORE = w_vec * norm(vec_score)
|
||||
+ w_time * time_decay
|
||||
+ w_graph * graph_context_score
|
||||
# Vorschlag: w_vec=0.70, w_time=0.20, w_graph=0.10; half_life_days=365
|
||||
```
|
||||
- Für Fragen, die „zeitlos“ sind, kann `half_life_days = inf` gesetzt werden (kein Decay).
|
||||
|
||||
### 16.4 Erklärbarkeit („Warum diese Antwort?“)
|
||||
- Jeder Antwort wird ein **Erklärpfad** mitgegeben (optional im UI):
|
||||
- Top-Chunks (mit Score-Zerlegung w_vec/w_time/w_graph)
|
||||
- **Kausalpfad** im Graphen: `Note → Edge(relation, confidence) → Note …`
|
||||
- **Provenienz** (Quelle, Erstellungszeit, letzte Änderung).
|
||||
- Ein **Multi-Layer-Explanations**-Modul ordnet Beiträge zu: Kindheit/Erlebnisse, Werte & Prinzipien, aktuelle Ziele/Leitbild, Gegenwartskontext.
|
||||
|
||||
### 16.5 Qdrant-Schemata (Ergänzungen, rückwärtskompatibel)
|
||||
- **Collections** bleiben: `mindnet_notes`, `mindnet_chunks`, `mindnet_edges` (+ `mindnet_people` falls noch nicht vorhanden).
|
||||
- Gemeinsame Zusatzfelder (alle Payloads):
|
||||
- `lang: "de"`
|
||||
- `visibility`/`sensitivity`/`acl` (s. 16.2)
|
||||
- `created_at` (ISO-8601), `updated_at` (ISO-8601), `age_days` (App-seitig berechnet)
|
||||
- **Chunks**:
|
||||
- Named-Vectors (z. B. `"content"`, optional `"title"`, `"tags"` für Hybrid-Scores).
|
||||
- `text_index` auf `body` für BM25-ähnliche Volltextsuche (Qdrant-Textindex).
|
||||
- `recency_group`: Jahr-Monat `YYYY-MM` (ermöglicht Pre-Filter für „letzte N Monate“).
|
||||
- **Edges** (explizit, keine Workarounds):
|
||||
- Felder: `src_id`, `dst_id`, `relation`, `direction`, `confidence` (0–1), `evidence` (Chunk-IDs), `created_at`, `updated_at`.
|
||||
- **edge_defaults** bleiben als **Interpretations-Hints**: z. B. `default_weight`, `symmetry`, `transitivity`. Sie ersetzen **nicht** echte Kanten, sondern parametrisieren abgeleitete/zusätzliche Kantenbildung.
|
||||
- Abgeleitete Edges: optionaler Batch-Job erzeugt *ergänzende* Kanten entlang `types.yaml`-Regeln (z. B. Backlinks, „inspired_by“ aus Zitaten), niemals statt der Originalkanten.
|
||||
|
||||
### 16.6 Regeln & Typen (konfigurierbar)
|
||||
- `types.yaml` erweitert um:
|
||||
- Relationstypen mit Default-Eigenschaften (`weight`, `directed`, `symmetric`, `transitive`).
|
||||
- **Scoring-Hooks**: pro Relation optionaler Einfluss auf `graph_context_score`.
|
||||
- **Privacy-Hooks**: pro Relation Minimal-Visibility (z. B. Beziehungen zu Personen sind mindestens `internal`).
|
||||
|
||||
### 16.7 Speicher- & Betriebsstrategie (heute lokal, später skalierbar)
|
||||
- **Heute**: Single-Node Qdrant mit HNSW-Index, Text-Index und Payload-Filtern; regelmäßiges Backup (Snapshots).
|
||||
- **Morgen**: Migrationspfad in **Qdrant-Distributed** (Shards, ggf. Replikation) bei Bedarf; Parameter-Gleichheit sichern (named-vectors, Metrik, HNSW-Param).
|
||||
- **Sicherheit**: Zugriff nur über internes Netz/Reverse-Proxy, Auth vor Qdrant; Festplatten-Verschlüsselung auf OS-Ebene (z. B. LUKS).
|
||||
|
||||
### 16.8 Persona-Layer („meine Essenz“)
|
||||
- Eigenständiger **Persona-Knoten** (z. B. Note `person/lars`), verbunden mit:
|
||||
- Werten/Prinzipien, Leitbild, prägenden Erlebnissen, Zielen, Rollen.
|
||||
- Jede Antwort kann gegen diesen Knoten **kontextualisiert** werden (Graph-Pfadnähe).
|
||||
- Feld `persona_weight` in Chunks/Notes zur Priorisierung persönlicher Relevanz.
|
||||
|
||||
### 16.9 Evaluierung & Qualitätssicherung
|
||||
- **Offline-Sets** aus echten Fragen (inkl. gewünschter Perspektiven) mit erwarteten Top-Quellen.
|
||||
- **Telemetrie** (lokal): Query-Arten, Treffer-Overlaps, Zeitgewicht-Effekte, „Warum-Pfad“ genutzt/ignoriert.
|
||||
- **Drift-Checks**: Anteil sehr alter vs. neuer Quellen in Antworten; manuelle Reviews.
|
||||
|
||||
### 16.10 Offene Punkte / Nächste Schritte
|
||||
1) `types.yaml` um Relationseigenschaften erweitern (s. 16.6).
|
||||
2) Retriever-Komposition inkl. Zeitgewicht (s. 16.3) implementieren.
|
||||
3) Explanations-Pfad inkl. Provenienzoberfläche (s. 16.4) im UI.
|
||||
4) Feld-Erweiterungen in Import-Pipelines (`lang`, `visibility`, Zeitfelder).
|
||||
5) Optional: Batch-Job für abgeleitete Edges nach Regeln (ergänzend, nicht ersetzend).
|
||||
@@ -0,0 +1,235 @@
|
||||
# Mindnet v2.2 – Fachliche Architektur
|
||||
**Datei:** `docs/mindnet_functional_architecture_v2.2.md`
|
||||
**Stand:** 2025-12-07
|
||||
**Status:** **FINAL** (Integrierter Stand WP01–WP04a)
|
||||
|
||||
> Dieses Dokument beschreibt **was** Mindnet fachlich tut und **warum** – mit Fokus auf die Erzeugung und Nutzung von **Edges** (Kanten) zwischen Notizen/Chunks sowie die Logik des Retrievers. Die technische Umsetzung (Module, Flags, Schemas) wird im technischen Dokument detailliert.
|
||||
|
||||
---
|
||||
|
||||
## 0) Zielbild & Grundprinzip
|
||||
|
||||
Mindnet wandelt Obsidian-Markdown-Notizen in einen **dynamischen Graphen aus Punkten und Kanten** um.
|
||||
Die drei zentralen Artefakt-Sammlungen lauten:
|
||||
|
||||
- `mindnet_notes` – genau **eine** Note pro Markdown-Datei (Metadaten & Hashes)
|
||||
- `mindnet_chunks` – semantische Teilstücke einer Note (Fenster/„Chunks“)
|
||||
- `mindnet_edges` – gerichtete Beziehungen zwischen Knoten (Chunks/Notes)
|
||||
|
||||
Die Import-Pipeline erzeugt diese Artefakte **deterministisch** und **idempotent** (erneute Läufe überschreiben konsistent statt zu duplizieren). Die Import-Schritte sind: *parse → chunk → edge → upsert*.
|
||||
|
||||
---
|
||||
|
||||
## 1) Notizen & Chunks (fachliche Perspektive)
|
||||
|
||||
### 1.1 Notiz (Note)
|
||||
- Repräsentiert eine fachliche Einheit (z. B. *„Vector DB Basics“*, *„KI-Projekt“*).
|
||||
- Trägt Eigenschaften (Titel, Typ, Zeitstempel, optionale Policies).
|
||||
- **Typ** (*type*) steuert u. a. Chunk-Profil und Default-Relationen (siehe §4).
|
||||
- **retriever_weight** und **chunk_profile** werden systematisch an Note **und** Chunks gespiegelt, damit der Retriever beides nutzen kann.
|
||||
|
||||
### 1.2 Chunk
|
||||
- Ausschnitt/Textfenster aus der Note, als eigenständiger Such-Anker.
|
||||
- Jeder Chunk gehört **genau einer** Note.
|
||||
- Chunks bilden eine Sequenz (1…N) – das ermöglicht *next/prev*.
|
||||
- **Neu in v2.2:** Alle Kanten entstehen ausschließlich zwischen Chunks (Scope="chunk"), nie zwischen Notes direkt. Notes dienen nur noch als Metadatencontainer.
|
||||
|
||||
> **Wichtig:** Chunking-Profile (short/medium/long) kommen aus `types.yaml` (per Note-Typ), können aber lokal überschrieben werden. Die effektiven Werte werden bei der Payload-Erzeugung bestimmt.
|
||||
|
||||
---
|
||||
|
||||
## 2) Edges – fachliche Typen & Bedeutung
|
||||
|
||||
Edges kodieren Beziehungen. Sie sind **gerichtet** und werden in `mindnet_edges` gespeichert. Die Erzeugung folgt einer strengen Priorität: **Inline > Callout > Defaults > Struktur**.
|
||||
|
||||
### 2.1 Struktur-Kanten (Das Skelett)
|
||||
- **belongs_to**: *Chunk → Note* (Zuordnung)
|
||||
- **next / prev**: Kettenbeziehungen zwischen aufeinanderfolgenden Chunks derselben Note.
|
||||
→ ermöglichen „lineares Weiterlesen“ oder Kontextpfade in Antworten.
|
||||
Diese Kanten entstehen immer, unabhängig von Inhalten.
|
||||
|
||||
### 2.2 Inhalts-Kanten (explizit)
|
||||
Hier unterscheidet v2.2 präzise zwischen verschiedenen Quellen der Evidenz:
|
||||
|
||||
1. **Explizite Inline-Relationen (Höchste Priorität):**
|
||||
Im Fließtext notierte, semantisch qualifizierte Relationen.
|
||||
* Syntax: `[[rel:depends_on Embeddings 101]]`
|
||||
* Fachliche Aussage: "Hängt ab von", "Ähnlich zu".
|
||||
* *rule_id*: `inline:rel`
|
||||
* *confidence*: ~0.95
|
||||
|
||||
2. **Callout-Edges (Kuratierte Listen):**
|
||||
Redaktionell gepflegte Listen am Ende einer Notiz (z.B. "Siehe auch").
|
||||
* Syntax: `> [!edge] related_to: [[Vector DB Basics]]`
|
||||
* *rule_id*: `callout:edge`
|
||||
* *confidence*: ~0.90
|
||||
|
||||
3. **Wikilinks (Standard-Referenzen):**
|
||||
Der klassische Obsidian-Link.
|
||||
* Syntax: `[[Vector DB Basics]]`
|
||||
* Typ: `references`
|
||||
* *rule_id*: `explicit:wikilink`
|
||||
* *confidence*: 1.0
|
||||
|
||||
> Alle expliziten Quellen (Wikilink, Inline-Rel, Callout) sind **Primärbelege**.
|
||||
|
||||
### 2.3 Typ-basierte Default-Kanten (Regelbasiert)
|
||||
Jede Note hat einen *Typ* (z. B. `concept`, `project`, `profile`). In `config/types.yaml` definieren wir pro Typ **edge_defaults**, z. B.:
|
||||
|
||||
- `concept`: `["references","related_to"]`
|
||||
- `project`: `["references","depends_on"]`
|
||||
|
||||
Regel: **Für jede gefundene explizite Referenz** (s. o.) werden **zusätzliche** Edges nach diesen Defaults erzeugt.
|
||||
Beispiel: Ein *project* mit `edge_defaults=["depends_on"]` erzeugt zu *jedem* explizit referenzierten Ziel **zusätzlich** eine `depends_on`-Kante.
|
||||
Diese Kanten tragen *provenance=rule* und eine **rule_id** der Form `edge_defaults:{note_type}:{relation}` sowie eine geringere Confidence (~0.7).
|
||||
|
||||
---
|
||||
|
||||
## 3) Edge-Payload – Felder & Semantik
|
||||
|
||||
Jede Kante hat mindestens:
|
||||
|
||||
- `kind` – Beziehungsart *(belongs_to, next, prev, references, related_to, depends_on, similar_to, …)*
|
||||
- `scope` – `"chunk"` (Standard in v2.2)
|
||||
- `source_id`, `target_id` – Quell-/Ziel-Knoten (Chunk-IDs oder Note-Titel bei unresolved Targets)
|
||||
- `note_id` – **Owner-Note** (die Note, aus der die Kante stammt)
|
||||
|
||||
Erweiterte/abgeleitete Felder (WP03 Superset):
|
||||
|
||||
- `provenance` – `"explicit"` (Wikilink/Inline/Callout) oder `"rule"` (Typ-Defaults)
|
||||
- `rule_id` – maschinenlesbare Regelquelle (z.B. `inline:rel`, `edge_defaults:project:depends_on`)
|
||||
- `confidence` – 0.0–1.0; Heuristik zur Gewichtung im Scoring.
|
||||
|
||||
**Dedup-Schlüssel** (fachlich): Gleichlautende `(source_id, target_id, kind, scope, rule_id)` werden nicht mehrfach geschrieben.
|
||||
|
||||
---
|
||||
|
||||
## 4) Typ-Registry (`config/types.yaml`)
|
||||
|
||||
### 4.1 Zweck
|
||||
- Steuert **Chunking-Profile** (`short|medium|long`) **pro Typ**
|
||||
- Liefert **retriever_weight** (Note-/Chunk-Gewichtung im Ranking) **pro Typ**
|
||||
- Definiert **edge_defaults** je Typ (s. o.)
|
||||
|
||||
Der Importer lädt die Registry aus `MINDNET_TYPES_FILE` oder nutzt Fallbacks. **Frontmatter-Overrides** für Profile werden in v2.2 weitgehend ignoriert zugunsten einer zentralen Governance in der YAML-Datei.
|
||||
|
||||
### 4.2 Beispiel (auslieferungsnah)
|
||||
version: 1.0
|
||||
types:
|
||||
concept:
|
||||
chunk_profile: medium
|
||||
edge_defaults: ["references", "related_to"]
|
||||
retriever_weight: 0.60
|
||||
project:
|
||||
chunk_profile: long
|
||||
edge_defaults: ["references", "depends_on"]
|
||||
retriever_weight: 0.97
|
||||
|
||||
**Auflösung im Importer**
|
||||
- `effective_chunk_profile(note_type, registry)` → `"short|medium|long"|None`
|
||||
- `effective_retriever_weight(note_type, registry)` → `float|None`
|
||||
- Ergebnis wird in `note_payload` **und** `chunk_payloads` gespiegelt.
|
||||
|
||||
---
|
||||
|
||||
## 5) Der Retriever (Funktionaler Layer)
|
||||
|
||||
Der Retriever ist in v2.2 der zentrale Zugangspunkt zum Wissen. Er realisiert die **Hybride Suche**.
|
||||
|
||||
### 5.1 Scoring-Modell
|
||||
Die Relevanz eines Treffers ergibt sich nicht mehr nur aus Textähnlichkeit, sondern aus einer gewichteten Formel:
|
||||
|
||||
total_score =
|
||||
semantic_weight * semantic_score
|
||||
+ edge_weight * edge_bonus
|
||||
+ centrality_weight * centrality_bonus
|
||||
+ type_weight * retriever_weight
|
||||
|
||||
* **semantic_score:** Vektorähnlichkeit (Qdrant).
|
||||
* **retriever_weight:** Wichtigkeit des Typs (z.B. Projekte > Quellen).
|
||||
* **edge_bonus:** Belohnung für Chunks, die im Kontext der Anfrage stark vernetzt sind (unter Berücksichtigung der *confidence*).
|
||||
* **centrality_bonus:** Belohnung für Knoten, die zentrale Hubs im lokalen Graphen darstellen.
|
||||
|
||||
### 5.2 Erklärbarkeit (Explainability)
|
||||
Der Retriever liefert **Begründungspfade**. Er gibt nicht nur Text zurück, sondern:
|
||||
* `score_breakdown`: Warum ist dieser Chunk oben? (War es der Text oder der Graph?)
|
||||
* `paths`: Über welche Kanten wurde dieser Chunk gefunden? (z.B. "Gefunden via `depends_on` von deiner aktuellen Projekt-Notiz").
|
||||
|
||||
---
|
||||
|
||||
## 6) Confidence & Provenance – wozu?
|
||||
|
||||
Der Retriever kann Edges gewichten:
|
||||
- **provenance**: *explicit* > *rule*
|
||||
- **confidence**: numerische Feinsteuerung
|
||||
- **retriever_weight (Note/Chunk)**: skaliert die Relevanz des gesamten Knotens
|
||||
|
||||
Eine typische Gewichtung (konfigurierbar in `retriever.yaml`) ist:
|
||||
`explicit: 1.0`, `rule: 0.8`. Damit bevorzugt der Graph **kuratiertes Wissen** (explizit notierte Links) vor „erweiterten“ Default-Ableitungen.
|
||||
|
||||
---
|
||||
|
||||
## 7) Semantik ausgewählter `kind`-Werte
|
||||
|
||||
- `references` – „Nutzt/erwähnt“; neutral, aber stützend für Kontext.
|
||||
- `related_to` – Ähnlichkeit/Verwandtschaft (symmetrisch interpretierbar).
|
||||
- `similar_to` – noch engere Ähnlichkeit; oft aus Inline-Rel (bewusst gesetzt).
|
||||
- `depends_on` – fachliche Abhängigkeit (z. B. „Projekt X hängt von Y ab“).
|
||||
- `belongs_to`, `next`, `prev` – Struktur.
|
||||
|
||||
> Symmetrische Relationen (z. B. `related_to`, `similar_to`) können **explizit** nur einseitig notiert sein, aber im Retriever beidseitig interpretiert werden.
|
||||
|
||||
---
|
||||
|
||||
## 8) Frontmatter-Eigenschaften – Rolle & Empfehlung
|
||||
|
||||
Frontmatter-Eigenschaften (Properties) bleiben **minimiert**:
|
||||
- **type** – steuert Typ-Defaults via Registry (Pflicht für differenziertes Verhalten).
|
||||
- **tags** – für klassische Filterung.
|
||||
- **retriever_weight / chunk_profile** – Sollten **nicht** manuell gesetzt werden (Best Practice: `types.yaml` nutzen).
|
||||
- **links** – Veraltet. Bitte Links primär über Text (Wikilink/Inline/Callout) pflegen.
|
||||
|
||||
---
|
||||
|
||||
## 9) Lösch-/Update-Garantien (Idempotenz)
|
||||
|
||||
- Jede Note hat einen stabilen **note_id** (Frontmatter/Hash).
|
||||
- Vor einem Upsert können *alte Chunks/Edges einer Note* gefiltert gelöscht werden (`note_id`-Filter) – das hält Collections sauber bei Re-Imports.
|
||||
- Die Import-Skripte unterstützen Modi wie `--purge-before-upsert` und `--sync-deletes`, um den Graph sauber zu halten.
|
||||
|
||||
---
|
||||
|
||||
## 10) Beispiel – Von Markdown zu Kanten (v2.2)
|
||||
|
||||
**Markdown (Auszug)**
|
||||
# Relations Showcase
|
||||
…
|
||||
Wir nutzen [[rel:depends_on Qdrant Vektordatenbank]] für die Suche.
|
||||
…
|
||||
Siehe auch: [[Embeddings 101]]
|
||||
|
||||
> [!edge] related_to: [[Vector DB Basics]]
|
||||
|
||||
**Ergebnis (fachlich)**
|
||||
1. `depends_on(Chunk→Qdrant)` mit `rule_id=inline:rel`, `confidence≈0.95`.
|
||||
2. `references(Chunk→Embeddings 101)` mit `rule_id=explicit:wikilink`, `confidence=1.0`.
|
||||
3. `related_to(Chunk→Vector DB Basics)` via Callout; `rule_id=callout:edge`, `confidence≈0.90`.
|
||||
4. **Typ-Defaults:** Falls die Note vom Typ `project` ist, entstehen zusätzlich `depends_on`-Kanten zu den Zielen aus (2) und (3).
|
||||
|
||||
---
|
||||
|
||||
## 11) Qualitäts- und Testkriterien (fachlich)
|
||||
|
||||
- **Keine Duplikate** gleicher `(src, relation, dst)` in der Collectionsicht (Dedup).
|
||||
- **Explizite Links** werden vollständig materialisiert (*explicit_total*).
|
||||
- **Typ-Defaults** ergänzen, aber überdecken nicht.
|
||||
- **Retriever Smoke-Test:** Ein Query muss plausible Scores liefern, die den Einfluss von `edge_bonus` zeigen.
|
||||
|
||||
---
|
||||
|
||||
## 12) Referenzen (Projektdateien & Leitlinien)
|
||||
|
||||
- Import-Pipeline & Registry-Auflösung: `scripts/import_markdown.py`.
|
||||
- Kantenbildung (V2-Logic): `app/core/derive_edges.py`.
|
||||
- Typ-Registry: `config/types.yaml` & `TYPE_REGISTRY_MANUAL.md`.
|
||||
- Retriever-Scoring: `app/core/retriever.py` & `wp04_retriever_scoring.md`.
|
||||
@@ -0,0 +1,215 @@
|
||||
# Mindnet v2.2 – Technische Architektur
|
||||
**Datei:** `docs/mindnet_technical_architecture_v2.2.md`
|
||||
**Stand:** 2025-12-07
|
||||
**Status:** **FINAL** (Integrierter Stand WP01–WP04a)
|
||||
**Quellen:** `mindnet_technical_architecture.md`, `chunking_strategy.md`, `Handbuch.md`, `wp04_retriever_scoring.md`, `Überarbeitungshinweise_WP03.md`, `Überarbeitungshinweise_WP04.md`.
|
||||
|
||||
> **Ziel dieses Dokuments:**
|
||||
> Vollständige, konsolidierte Beschreibung der aktuellen technischen Architektur von **Mindnet v2.2**. Es definiert die Datenstrukturen in Qdrant, die Verarbeitungspipelines (Importer, Chunker, Edges) und die Retrieval-Logik. Es bildet den **aktuellen Implementierungsstand** ab.
|
||||
|
||||
---
|
||||
|
||||
## 1. Systemüberblick
|
||||
|
||||
### 1.1 Architektur-Zielbild
|
||||
Mindnet ist ein **persönliches Wissensnetz**. Technisch bedeutet das:
|
||||
1. **Source:** Markdown-Notizen in einem Vault (Obsidian-kompatibel).
|
||||
2. **Pipeline:** Ein Python-Importer transformiert diese in **Notes**, **Chunks** und **Edges**.
|
||||
3. **Storage:** Speicherung in **Qdrant** (Vektor-Datenbank) in drei getrennten Collections.
|
||||
4. **Service:** Eine FastAPI-Anwendung stellt Endpunkte für **Semantische** und **Hybride Suche** bereit.
|
||||
|
||||
Das System arbeitet **deterministisch** (stabile IDs) und ist **konfigurationsgetrieben** (`types.yaml`, `retriever.yaml`).
|
||||
|
||||
### 1.2 Verzeichnisstruktur & Komponenten
|
||||
Die Codebasis ist modular aufgebaut.
|
||||
|
||||
/mindnet/
|
||||
├── app/
|
||||
│ ├── main.py # FastAPI Einstiegspunkt
|
||||
│ ├── core/
|
||||
│ │ ├── qdrant.py # Client-Factory & Connection
|
||||
│ │ ├── qdrant_points.py # Low-Level Point Operations (Upsert/Delete)
|
||||
│ │ ├── note_payload.py # Bau der Note-Objekte
|
||||
│ │ ├── chunk_payload.py # Bau der Chunk-Objekte
|
||||
│ │ ├── chunker.py # Text-Zerlegung (Profiling)
|
||||
│ │ ├── edges.py # Edge-Datenstrukturen
|
||||
│ │ ├── derive_edges.py # Logik der Kantenableitung (WP03)
|
||||
│ │ └── retriever.py # Scoring & Graph-Expansion (WP04)
|
||||
│ ├── models/ # Pydantic DTOs (QueryRequest, etc.)
|
||||
│ └── routers/ # API Routen
|
||||
├── config/
|
||||
│ ├── types.yaml # Typ-Definitionen (Import-Zeit)
|
||||
│ └── retriever.yaml # Scoring-Gewichte (Laufzeit)
|
||||
├── scripts/
|
||||
│ ├── import_markdown.py # Haupt-Importer CLI
|
||||
│ ├── payload_dryrun.py # Diagnose: JSON-Generierung ohne DB
|
||||
│ └── edges_full_check.py # Diagnose: Graph-Integrität
|
||||
└── tests/ # Pytest Suite
|
||||
|
||||
---
|
||||
|
||||
## 2. Datenmodell & Collections (Qdrant)
|
||||
|
||||
Das Datenmodell verteilt sich auf drei Collections, definiert durch `COLLECTION_PREFIX` (Standard: `mindnet`).
|
||||
|
||||
### 2.1 Notes Collection (`<prefix>_notes`)
|
||||
Repräsentiert die Metadaten einer Datei.
|
||||
* **Zweck:** Filterung, Metadaten-Haltung, Vererbung von Eigenschaften an Chunks.
|
||||
* **Schema (Payload):**
|
||||
|
||||
| Feld | Datentyp | Beschreibung | Herkunft |
|
||||
| :--- | :--- | :--- | :--- |
|
||||
| `note_id` | Keyword | Stabile ID (UUIDv5 oder Slug). | Frontmatter |
|
||||
| `title` | Text | Titel der Notiz. | Frontmatter |
|
||||
| `type` | Keyword | Logischer Typ (z.B. `concept`). | `types.yaml` Resolver |
|
||||
| `retriever_weight` | Float | Wichtigkeit im Retrieval. | `types.yaml` |
|
||||
| `chunk_profile` | Keyword | Genutztes Chunking-Profil. | `types.yaml` |
|
||||
| `edge_defaults` | List[Str] | Aktive Default-Relationen. | `types.yaml` |
|
||||
| `tags` | List[Kw] | Tags zur Filterung. | Frontmatter |
|
||||
| `updated` | Integer | Zeitstempel (z.B. YYYYMMDD...). | File Stats |
|
||||
| `fulltext` | Text | Gesamter Inhalt (für Export/Rekonstruktion). | Body |
|
||||
|
||||
|
||||
|
||||
### 2.2 Chunks Collection (`<prefix>_chunks`)
|
||||
Die atomaren Sucheinheiten.
|
||||
* **Zweck:** Vektorsuche (Embeddings), Granulares Ergebnis.
|
||||
* **Schema (Payload):**
|
||||
|
||||
| Feld | Datentyp | Beschreibung |
|
||||
| :--- | :--- | :--- |
|
||||
| `chunk_id` | Keyword | Deterministisch: `{note_id}#c{index:02d}`. |
|
||||
| `note_id` | Keyword | Referenz zur Note. |
|
||||
| `text` | Text | **Reiner Inhalt** (ohne Overlap). Anzeige-Text. |
|
||||
| `window` | Text | **Kontext-Fenster** (mit Overlap). Embedding-Basis. |
|
||||
| `ord` | Integer | Sortierreihenfolge (1..N). |
|
||||
| `retriever_weight` | Float | Vererbt von Note. |
|
||||
| `chunk_profile` | Keyword | Vererbt von Note. |
|
||||
| `neighbors_prev` | List[Str] | ID des Vorgänger-Chunks. |
|
||||
| `neighbors_next` | List[Str] | ID des Nachfolger-Chunks. |
|
||||
|
||||
|
||||
|
||||
### 2.3 Edges Collection (`<prefix>_edges`)
|
||||
Gerichtete Kanten. Massiv erweitert in WP03 für Provenienz-Tracking.
|
||||
* **Zweck:** Graph-Traversal, Kontext-Anreicherung.
|
||||
* **Schema (Payload):**
|
||||
|
||||
| Feld | Datentyp | Beschreibung | Wertebereich (Bsp.) |
|
||||
| :--- | :--- | :--- | :--- |
|
||||
| `edge_id` | Keyword | Hash aus (src, dst, kind). | |
|
||||
| `source_id` | Keyword | ID des Start-Chunks. | |
|
||||
| `target_id` | Keyword | ID des Ziel-Chunks (oder Note-Titel). | |
|
||||
| `note_id` | Keyword | Note, die diese Kante definiert. | |
|
||||
| `kind` | Keyword | Art der Beziehung. | `references`, `depends_on`, `next` |
|
||||
| `scope` | Keyword | Geltungsbereich. | Immer `"chunk"` (v2.2). |
|
||||
| `rule_id` | Keyword | Herkunftsregel. | `explicit:wikilink`, `inline:rel` |
|
||||
| `confidence` | Float | Vertrauenswürdigkeit (0.0-1.0). | 1.0, 0.95, 0.7 |
|
||||
|
||||
|
||||
|
||||
---
|
||||
|
||||
## 3. Konfiguration
|
||||
|
||||
Die Logik ist ausgelagert in YAML-Dateien.
|
||||
|
||||
### 3.1 Typ-Registry (`config/types.yaml`)
|
||||
Steuert den Import-Prozess.
|
||||
* **Priorität:** Frontmatter > Pfad > Default.
|
||||
* **Inhalt:**
|
||||
```yaml
|
||||
types:
|
||||
concept:
|
||||
chunk_profile: medium
|
||||
edge_defaults: ["references", "related_to"]
|
||||
retriever_weight: 0.60
|
||||
```
|
||||
|
||||
### 3.2 Retriever-Config (`config/retriever.yaml`)
|
||||
Steuert das Scoring zur Laufzeit (WP04a).
|
||||
* **Inhalt:**
|
||||
```yaml
|
||||
scoring:
|
||||
semantic_weight: 1.0
|
||||
edge_weight: 0.7
|
||||
centrality_weight: 0.5
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 4. Import-Pipeline (Markdown → Qdrant)
|
||||
|
||||
Das Skript `scripts/import_markdown.py` orchestriert den Prozess.
|
||||
|
||||
### 4.1 Verarbeitungsschritte
|
||||
|
||||
1. **Discovery & Parsing:**
|
||||
* Einlesen der `.md` Dateien. Hash-Vergleich (Body/Frontmatter) zur Erkennung von Änderungen.
|
||||
2. **Typauflösung:**
|
||||
* Laden der `types.yaml`. Bestimmen des effektiven Typs und der `edge_defaults`.
|
||||
3. **Chunking:**
|
||||
* Zerlegung via `chunker.py` basierend auf `chunk_profile` (z.B. `by_heading`, `short`, `long`).
|
||||
* Trennung von `text` (Kern) und `window` (Embedding-Kontext).
|
||||
4. **Kantenableitung (Edge Derivation):**
|
||||
Die `derive_edges.py` erzeugt Kanten in strikter Reihenfolge:
|
||||
1. **Inline-Edges:** `[[rel:depends_on X]]` → `kind=depends_on`, `rule_id=inline:rel`, `conf=0.95`.
|
||||
2. **Callout-Edges:** `> [!edge] related_to: [[X]]` → `kind=related_to`, `rule_id=callout:edge`, `conf=0.90`.
|
||||
3. **Explizite Referenzen:** `[[X]]` → `kind=references`, `rule_id=explicit:wikilink`, `conf=1.0`.
|
||||
4. **Typ-Defaults:** Für jede Referenz werden Zusatzkanten gemäß `edge_defaults` erzeugt (z.B. `project` -> `depends_on`). `rule_id=edge_defaults:...`, `conf=0.7`.
|
||||
5. **Struktur:** `belongs_to`, `next`, `prev` (automatisch).
|
||||
5. **Upsert:**
|
||||
* Schreiben in Qdrant. Nutzung von `--purge-before-upsert` für saubere Updates.
|
||||
|
||||
---
|
||||
|
||||
## 5. Retriever-Architektur & Scoring
|
||||
|
||||
Der Retriever (`app/core/retriever.py`) realisiert die Logik hinter `/query`.
|
||||
|
||||
### 5.1 Betriebsmodi
|
||||
* **Semantic:** Reine Vektorsuche. Schnell.
|
||||
* **Hybrid:** Vektorsuche + Graph-Expansion (Tiefe N) + Re-Ranking.
|
||||
|
||||
### 5.2 Scoring-Formel (WP04a)
|
||||
Die Relevanzberechnung ist nun eine gewichtete Summe:
|
||||
|
||||
$$
|
||||
TotalScore = (W_{sem} \cdot S_{sem} \cdot \max(W_{type}, 0)) + (W_{edge} \cdot B_{edge}) + (W_{cent} \cdot B_{cent})
|
||||
$$
|
||||
|
||||
* **Komponenten:**
|
||||
* $S_{sem}$: Semantic Score (Cosine Similarity).
|
||||
* $W_{type}$: `retriever_weight` (aus Note/Chunk Payload).
|
||||
* $B_{edge}$: Edge-Bonus (Summe der Konfidenzen eingehender relevanter Kanten).
|
||||
* $B_{cent}$: Centrality-Bonus (im lokalen Subgraphen).
|
||||
* **Gewichte ($W$):** Stammen aus `retriever.yaml`.
|
||||
|
||||
### 5.3 Graph-Expansion
|
||||
Der Hybrid-Modus lädt dynamisch die Nachbarschaft der Top-K Vektor-Treffer ("Seeds") über `graph_adapter.expand`. Dies baut einen temporären `NetworkX`-Graphen im Speicher, auf dem Boni berechnet werden.
|
||||
|
||||
---
|
||||
|
||||
## 6. Indizes & Performance
|
||||
|
||||
Damit Qdrant performant bleibt, sind Payload-Indizes essenziell.
|
||||
|
||||
**Erforderliche Indizes:**
|
||||
* **Notes:** `note_id`, `type`, `tags`.
|
||||
* **Chunks:** `note_id`, `chunk_id`.
|
||||
* **Edges:** `source_id`, `target_id`, `kind`, `scope`, `note_id`.
|
||||
|
||||
Validierung erfolgt über `tests/ensure_indexes_and_show.py`.
|
||||
|
||||
---
|
||||
|
||||
## 7. Offene Punkte / Known Limitations
|
||||
|
||||
1. **Multi-Target Inline-Relations:**
|
||||
* `rel: similar_to [[A]] [[B]]` wird aktuell parser-seitig nicht unterstützt.
|
||||
* Workaround: Zwei separate Inline-Links `[[rel:similar_to A]] [[rel:similar_to B]]`.
|
||||
2. **Unresolved Targets:**
|
||||
* Kanten zu Notizen, die noch nicht existieren, werden mit `target_id="Titel"` angelegt.
|
||||
* Heilung durch `scripts/resolve_unresolved_references.py` möglich.
|
||||
3. **Vektor-Konfiguration für Edges:**
|
||||
* `mindnet_edges` hat aktuell keine Vektoren (`vectors = null`). Eine semantische Suche *auf Kanten* ist noch nicht möglich.
|
||||
@@ -0,0 +1,365 @@
|
||||
# mindnet v2 – Implementierungs-Playbook & Handover (Side‑by‑Side zu v1)
|
||||
*Stand:* 2025-11-10 08:03
|
||||
*Autor:* ChatGPT (Projekt‑Handover)
|
||||
*Zweck:* Vollständige, ausführbare Anleitung zur Weiterentwicklung von **mindnet** auf eine versionierte, testbare v2‑Architektur (neben v1), inklusive Prompt für den Folgethread, Akzeptanzkriterien, Dateiliste, Test‑ und Rollback‑Pläne.
|
||||
|
||||
---
|
||||
|
||||
## 0) Kontext & Zielbild (kurz)
|
||||
- **Mission:** Persönliches Wissensnetz, das langfristig die eigene Persönlichkeit, Erfahrungen und Entscheidungslogik abbildet; später nachvollziehbare Erklärungen für Familie/Nachwelt.
|
||||
- **Heutiger Stand (v1):** `mindnet_notes` in Qdrant vorhanden; **Chunks** und **Edges** bisher nur teilweise/nicht konsistent; Edge‑Defaults („1 Dot“) vorhanden, aber Ziele teils nicht materialisiert. Importskripte: `import_markdown.py`, `edges.py` (historisch).
|
||||
- **Ziel (v2):**
|
||||
- **Collections:** `mindnet_notes_v2`, `mindnet_chunks_v2`, `mindnet_edges_v2` (saubere Payload‑Schemata + Indizes).
|
||||
- **Chunker v2:** Block/Heading‑aware, Ziel 900–1200 Zeichen, Überlappung 120–180, deterministische `chunk_id`s.
|
||||
- **Edge‑Builder v2:** Reihenfolge: **explicit → rule → default_resolved → default**; saubere `provenance`.
|
||||
- **Importer‑Pipeline v2:** Parse → Chunk → Edge → Embeddings → Batch‑Upsert + Payload‑Indexpflege + Snapshots.
|
||||
- **Policies:** Privacy/Recency/Type‑Gewichte als **Konfiguration** (YAML + JSON‑Schema), nicht hart im Code.
|
||||
- **Observability:** FastAPI + OpenTelemetry (Tracing/Metrics); Import‑Reports (CSV/JSON).
|
||||
- **Side‑by‑Side:** v2 **parallel** zu v1; kein Big‑Bang.
|
||||
|
||||
> **Warum so?** Qdrant‑Payload‑Indizes/Filter → schnelle, erklärbare Selektionen; Snapshots → betriebssicheres Rollback; OTel → Nachvollziehbarkeit; YAML/JSON‑Schema → Validierbarkeit & Stabilität.
|
||||
|
||||
### 0.1 Aktueller Implementierungsstand (2025-11-18)
|
||||
|
||||
- Es existiert eine produktive mindnet-Instanz mit den Collections:
|
||||
- `${prefix}_notes`
|
||||
- `${prefix}_chunks`
|
||||
- `${prefix}_edges`
|
||||
- Import-Pipeline:
|
||||
- `scripts/import_markdown.py` erzeugt Notes, Chunks und Edges vollständig.
|
||||
- `types.yaml` definiert per Note-Typ:
|
||||
- `retriever_weight`
|
||||
- `chunk_profile`
|
||||
- `edge_defaults` (z. B. `depends_on`, `related_to`).
|
||||
- Edges:
|
||||
- Strukturkanten: `belongs_to`, `next`, `prev`
|
||||
- Explizite Kanten aus Wikilinks: `kind="references"`, `rule_id="explicit:wikilink"`
|
||||
- Typbasierte Default-Kanten aus `edge_defaults`: `rule_id="edge_defaults:<type>:<relation>"`
|
||||
- Inline-Kanten im Format `[[rel:<relation> <Zielnote>]]` mit `rule_id="inline:rel"`
|
||||
- alle Edges tragen eine `confidence` im Bereich `[0,1]` als Hinweis für spätere Ranking-Logiken.
|
||||
- Tests:
|
||||
- `tests/test_edges_smoke.py`, `tests/test_edges_all.py` und `scripts/edges_full_check.py` validieren Counts, Konsistenz (`belongs_to == chunks`, `next/prev`-Ketten, keine Duplikate) und Provenienz-Zählungen (`explicit`, `defaults`, `inline`, `structure`).
|
||||
- Offene Punkte:
|
||||
- Mehrere Ziele in einer Inline-Relation (`rel: similar_to [[A]] [[B]]`) sind noch nicht implementiert.
|
||||
- Retrieval-Service nutzt die neuen Edges und Weights bisher nur teilweise; ein edge-bewusster Retriever ist als nächster Schritt geplant.
|
||||
|
||||
|
||||
---
|
||||
|
||||
## 1) Datenmodelle (Schemata, v2)
|
||||
Die tatsächlichen JSON‑Schemata (2020‑12) werden im Repo abgelegt und in CI validiert.
|
||||
|
||||
### 1.1 note.schema.json
|
||||
```json
|
||||
{
|
||||
"$schema": "https://json-schema.org/draft/2020-12/schema",
|
||||
"$id": "https://example.org/mindnet/note.schema.json",
|
||||
"type": "object",
|
||||
"required": ["id", "title", "type", "privacy", "created", "hash_body"],
|
||||
"properties": {
|
||||
"id": { "type": "string", "pattern": "^[a-zA-Z0-9_-]+$" },
|
||||
"title": { "type": "string", "minLength": 1 },
|
||||
"type": { "type": "string" },
|
||||
"privacy": { "type": "string", "enum": ["public","internal","private"] },
|
||||
"created": { "type": "string", "format": "date-time" },
|
||||
"modified": { "type": "string", "format": "date-time" },
|
||||
"tags": { "type": "array", "items": { "type": "string" } },
|
||||
"lang": { "type": "string", "default": "de" },
|
||||
"source_path": { "type": "string" },
|
||||
"source_collection": { "type": "string" },
|
||||
"hash_body": { "type": "string" },
|
||||
"hash_frontmatter": { "type": "string" },
|
||||
"token_count": { "type": "integer", "minimum": 0 }
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
### 1.2 chunk.schema.json
|
||||
```json
|
||||
{
|
||||
"$schema": "https://json-schema.org/draft/2020-12/schema",
|
||||
"$id": "https://example.org/mindnet/chunk.schema.json",
|
||||
"type": "object",
|
||||
"required": ["chunk_id","note_id","text","ord"],
|
||||
"properties": {
|
||||
"chunk_id": { "type": "string", "pattern": "^[a-zA-Z0-9_-]+#c\d{4}$" },
|
||||
"note_id": { "type": "string" },
|
||||
"text": { "type": "string" },
|
||||
"ord": { "type": "integer", "minimum": 0 },
|
||||
"span_char_start": { "type": "integer", "minimum": 0 },
|
||||
"span_char_end": { "type": "integer", "minimum": 0 },
|
||||
"heading_path": { "type": "array", "items": { "type": "string" } },
|
||||
"section_title": { "type": "string" },
|
||||
"tokens_start": { "type": "integer", "minimum": 0 },
|
||||
"tokens_end": { "type": "integer", "minimum": 0 },
|
||||
"embeddings_version": { "type": "string" }
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
### 1.3 edge.schema.json
|
||||
```json
|
||||
{
|
||||
"$schema": "https://json-schema.org/draft/2020-12/schema",
|
||||
"$id": "https://example.org/mindnet/edge.schema.json",
|
||||
"type": "object",
|
||||
"required": ["edge_id","src_note_id","relation","provenance"],
|
||||
"properties": {
|
||||
"edge_id": { "type": "string" },
|
||||
"src_note_id": { "type": "string" },
|
||||
"src_chunk_id": { "type": "string" },
|
||||
"dst_note_id": { "type": "string" },
|
||||
"dst_chunk_id": { "type": "string" },
|
||||
"relation": { "type": "string" },
|
||||
"evidence_spans":{ "type": "array", "items": {"type":"object","properties":{"chunk_id":{"type":"string"},"span":[{"type":"integer"},{"type":"integer"}]]} },
|
||||
"provenance": { "type": "string", "enum": ["explicit","rule","default_resolved","default"] },
|
||||
"rule_id": { "type": "string" },
|
||||
"confidence": { "type": "number", "minimum": 0, "maximum": 1 }
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
### 1.4 default_edge.schema.json
|
||||
```json
|
||||
{
|
||||
"$schema": "https://json-schema.org/draft/2020-12/schema",
|
||||
"$id": "https://example.org/mindnet/default_edge.schema.json",
|
||||
"type": "object",
|
||||
"required": ["src_note_id","relation","target_kind"],
|
||||
"properties": {
|
||||
"src_note_id": { "type": "string" },
|
||||
"relation": { "type": "string" },
|
||||
"target_kind": { "type": "string" },
|
||||
"when": { "type": "string" },
|
||||
"strength_hint":{ "type": "number", "minimum": 0, "maximum": 1 }
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 2) Qdrant: Collections & Indizes (v2)
|
||||
Anlegen **neben** v1:
|
||||
- `mindnet_notes_v2` (Payload‑Index: `type, privacy, created, modified, tags, source_path`)
|
||||
- `mindnet_chunks_v2` (Payload‑Index: `note_id, ord, heading_path`)
|
||||
- `mindnet_edges_v2` (Payload‑Index: `src_note_id, dst_note_id, relation, provenance`)
|
||||
|
||||
> Snapshots für Backups/Restore einplanen (Runbook siehe unten).
|
||||
|
||||
---
|
||||
|
||||
## 3) Import‑Pipeline v2 – Module & Flags
|
||||
- **Chunker v2** (`chunking/block_chunker.py`): `target_len=1000`, `overlap=150`, `respect_headings=True`, `min_chunk_len=600`, `max_chunk_len=1400`, `chunk_id = f\"{note_id}#c{ordinal:04d}\"`
|
||||
- **Edge‑Builder v2** (`graph/edge_builder_v2.py`): Reihenfolge **explicit → rule → default_resolved → default**, `provenance` korrekt setzen.
|
||||
- **Importer** (`scripts/import_markdown.py` erweitern oder `scripts/import_markdown_v2.py`):
|
||||
Flags: `--schema v2`, `--chunker v2`, `--edges v2`, `--dry-run`, `--apply`, `--prefix "$COLLECTION_PREFIX"`.
|
||||
|
||||
**Beispiel‑Run (dry‑run):**
|
||||
```bash
|
||||
python3 -m scripts.import_markdown_v2 --vault ./vault --schema v2 --chunker v2 --edges v2 --dry-run --prefix "$COLLECTION_PREFIX"
|
||||
```
|
||||
|
||||
**Apply:**
|
||||
```bash
|
||||
python3 -m scripts.import_markdown_v2 --vault ./vault --schema v2 --chunker v2 --edges v2 --apply --prefix "$COLLECTION_PREFIX"
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 4) Policies (konfigurierbar, kein Hard‑Code)
|
||||
- `policies/retrieval.schema.json` (JSON‑Schema) und `policies/retrieval.yaml`, z. B.:
|
||||
```yaml
|
||||
privacy_order: [private, internal, public]
|
||||
recency_boost_half_life_days: 90
|
||||
type_priority:
|
||||
person: 1.2
|
||||
event: 1.1
|
||||
concept: 1.0
|
||||
max_chunks_per_note: 8
|
||||
edge_provenance_weights:
|
||||
explicit: 1.0
|
||||
rule: 0.8
|
||||
default_resolved: 0.6
|
||||
default: 0.2
|
||||
rule_sets:
|
||||
- id: "rules:types-v1"
|
||||
enabled: true
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 5) Observability & Reports
|
||||
- **FastAPI + OpenTelemetry**: Traces um Phasen *parse/chunk/edge/upsert*; Metriken (Counter/Histogramme) pro Phase.
|
||||
- **Import‑Report** je Lauf (CSV/JSON): counts für Notes/Chunks/Edges, Fehler, Dauer, Embedding‑Version.
|
||||
|
||||
---
|
||||
|
||||
## 6) Teststrategie
|
||||
- **Gold‑Notizen** (3–5 repräsentative .md) mit erwarteten *Chunk‑Counts* und *Edge‑Sets*.
|
||||
- **Regression:** fixer Soll‑Count (z. B. 171 ± 15 %) über Gesamt‑Vault; keine Duplikate `(src,relation,dst)`.
|
||||
- **Dry‑Run vs Apply**: identische Counts, nur ohne Upserts.
|
||||
- **Filter‑Smoke‑Tests:** Qdrant‑Filter nach `privacy`, `type`, `tags`, `created` funktionieren performant.
|
||||
|
||||
---
|
||||
|
||||
## 7) Akzeptanzkriterien (KPI)
|
||||
1. **Chunk‑Qualität:** ≥ 90 % der Chunks enden an semantischen Grenzen; Ziel‑Länge 900–1200 Zeichen; deterministische IDs.
|
||||
2. **Edge‑Kohärenz:** Keine Duplikate gleicher `(src,relation,dst)`; ≥ 95 % der expliziten Links materialisiert; `default` nur, wenn kein Ziel existiert.
|
||||
3. **Filterbarkeit:** Queries nach `privacy/type/tags/created` performant (Payload‑Index vorhanden).
|
||||
4. **Betriebsfestigkeit:** Telemetrie aktiv; fehlertoleranter Import; Qdrant‑Snapshot nach erfolgreichem Lauf.
|
||||
|
||||
---
|
||||
|
||||
## 8) Roadmap in kleinen, testbaren Schritten
|
||||
|
||||
### Step 0 – Safeguards
|
||||
- Qdrant‑Snapshot v1 erstellen + Restore‑Probe.
|
||||
- ENV: `MINDNET_SCHEMA_VERSION=2` (Importer wird v2 schreiben, v1 bleibt unberührt).
|
||||
|
||||
### Step 1 – Schemata (nur Dateien)
|
||||
- Ablage: `/schemas/note.schema.json`, `/schemas/chunk.schema.json`, `/schemas/edge.schema.json`, `/schemas/default_edge.schema.json`
|
||||
- CI‑Job: `make schema-validate` (jsonschema).
|
||||
- **Abnahme:** Validator grün auf 3 Gold‑Notizen.
|
||||
|
||||
### Step 2 – Qdrant v2‑Collections (Optional)
|
||||
- Anlegen der 3 v2‑Collections + Payload‑Indizes.
|
||||
- **Abnahme:** Filter‑Query liefert erwartbare Ergebnisse.
|
||||
|
||||
### Step 2a – Edge-Pipeline v1 stabilisieren (ERLEDIGT)
|
||||
- Ziel: Vollständige Generierung von Struktur-, expliziten, typbasierten und Inline-Kanten in den bestehenden Collections `${prefix}_notes`, `${prefix}_chunks`, `${prefix}_edges`.
|
||||
- Status: Implementiert und durch `tests/test_edges_smoke.py`, `tests/test_edges_all.py` und `scripts/edges_full_check.py` verifiziert.
|
||||
|
||||
|
||||
### Step 3 – Chunker v2
|
||||
- Implementierung & Flag `--chunker v2`.
|
||||
- **Abnahme:** Chunk‑Counts ~ alt (≈171 ± 15 %), semantische Schnitte.
|
||||
|
||||
### Step 4 – Edge‑Builder v2
|
||||
- Reihenfolge & `provenance` strikt umsetzen, `default_resolved` integrieren.
|
||||
- **Abnahme:** erwartete Relationensätze auf Gold‑Notizen, keine Duplikate.
|
||||
|
||||
### Step 4a – Edge-bewusster Retriever
|
||||
- Implementiere einen Retrieval-Service, der:
|
||||
- Chunks über Vektor-Suche aus `${prefix}_chunks` holt,
|
||||
- dazu passende Nachbar-Chunks (`next/prev`) und typrelevante Notizen (`edge_defaults`, Inline-Relationen, Wikilinks) über `${prefix}_edges` ergänzt,
|
||||
- `retriever_weight` und `confidence` beim Scoring berücksichtigt.
|
||||
- Liefere einen API-Endpoint `/mindnet/query` mit einem einfachen JSON-Response (Chunks + Begründung).
|
||||
- Akzeptanz:
|
||||
- mindestens ein Testfall mit dem Relations-Showcase-Vault,
|
||||
- sichtbarer Einfluss von `edge_defaults` und Inline-Relationen auf die Ergebnisliste.
|
||||
|
||||
|
||||
### Step 5 – Importer‑Pipeline v2
|
||||
- `--schema v2` Side‑by‑Side; Batch‑Upserts; vollständige Payloads; Recency‑Boost Konfig **nur** in Policy.
|
||||
- **Abnahme:** Dry‑Run/Apply‑Parität, Reports, Telemetrie‑Ereignisse sichtbar.
|
||||
|
||||
### Step 6 – Observability
|
||||
- OTel‑Instrumentierung + `/import/status` + Reports.
|
||||
- **Abnahme:** Traces & Metriken vorhanden.
|
||||
|
||||
### Step 7 – Snapshots & Runbook
|
||||
- Snapshot nach Erfolg; Restore‑Dokument.
|
||||
- **Abnahme:** Restore‑Probe erfolgreich.
|
||||
|
||||
### Step 8 – Policies
|
||||
- `policies/retrieval.yaml` + JSON‑Schema; A/B‑Test (Policy ändern → Retrieval ändert sich ohne Reimport).
|
||||
- **Abnahme:** Sichtbarer Effekt laut Testfall.
|
||||
|
||||
### Step 9 – Umschalten
|
||||
- Abnahme‑Doku, KPI‑Delta dokumentiert; Alias/Flag‑Switch auf v2.
|
||||
- **Abnahme:** Funktionale Gleichwertigkeit + Qualitätsgewinn.
|
||||
|
||||
---
|
||||
|
||||
## 9) Dateien (neu/ändern)
|
||||
**Neu:**
|
||||
- `schemas/note.schema.json`
|
||||
- `schemas/chunk.schema.json`
|
||||
- `schemas/edge.schema.json`
|
||||
- `schemas/default_edge.schema.json`
|
||||
- `policies/retrieval.schema.json`
|
||||
- `policies/retrieval.yaml`
|
||||
- `chunking/block_chunker.py`
|
||||
- `graph/edge_builder_v2.py`
|
||||
- `docs/OPERATIONS.md` (Snapshots/Restore/Runbook)
|
||||
- `tests/gold_notes/manifest.yaml` (+ 3–5 Notizen Kopien)
|
||||
|
||||
**Änderungen:**
|
||||
- `scripts/import_markdown.py` **oder** `scripts/import_markdown_v2.py` neu (Flags `--schema/--chunker/--edges`).
|
||||
- `edges.py` (falls weiterverwendet): auf **v2‑Edge‑Builder** migrieren oder auslaufen lassen.
|
||||
- `types.yaml`: Regel‑Definitionen klarisieren (IDs, Bedingungen), **nicht** hart in Code.
|
||||
|
||||
---
|
||||
|
||||
## 10) Beispiel‑Kommandos
|
||||
```bash
|
||||
# Step 2: Collections prüfen
|
||||
curl -s http://127.0.0.1:6333/collections | jq
|
||||
|
||||
# Step 3: Chunker v2 (dry-run)
|
||||
python3 -m scripts.import_markdown_v2 --vault ./vault --schema v2 --chunker v2 --edges v2 --dry-run --prefix "$COLLECTION_PREFIX"
|
||||
|
||||
# Step 5: Apply + Report
|
||||
python3 -m scripts.import_markdown_v2 --vault ./vault --schema v2 --chunker v2 --edges v2 --apply --prefix "$COLLECTION_PREFIX" --report ./reports/import_$(date +%F_%H%M).json
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 11) Rollback
|
||||
- v2‑Collections droppen (wenn Side‑by‑Side).
|
||||
- Aus Snapshot wiederherstellen (v1).
|
||||
- Flags zurück auf v1.
|
||||
|
||||
---
|
||||
|
||||
## 12) Referenzen (offizielle Quellen)
|
||||
- Qdrant – Collections, Payload‑Index, Filter: <https://qdrant.tech/documentation/>
|
||||
- Qdrant – Snapshots/Backup: <https://qdrant.tech/documentation/guides/backup/>
|
||||
- OpenTelemetry (Python/FastAPI): <https://opentelemetry.io/docs/instrumentation/python/>
|
||||
- JSON‑Schema 2020‑12: <https://json-schema.org/>
|
||||
- YAML 1.2: <https://yaml.org/spec/1.2.2/>
|
||||
- Obsidian Frontmatter: <https://help.obsidian.md/Editing+and+formatting/Properties>
|
||||
- Ollama CLI/API (lokale Modelle): <https://github.com/ollama/ollama>
|
||||
|
||||
---
|
||||
|
||||
## 13) **PROMPT für den neuen Chat (bitte exakt so einfügen)**
|
||||
```
|
||||
Rolle: Du bist mein Senior‑Entwickler & Architekt für mindnet. Arbeite strikt in kleinen, testbaren Schritten (Side‑by‑Side v2 neben v1). Liefere komplette Dateien als Downloads. Frage immer zuerst nach den **aktuellsten** Projektdateien, wenn du sie brauchst (z. B. import_markdown.py, edges.py, types.yaml), ändere keine Systemfunktion „workaround‑artig“ ohne die Gesamtwirkung zu prüfen.
|
||||
|
||||
Kontext (Kurzfassung):
|
||||
- v1 hat nur mindnet_notes zuverlässig. Chunks/Edges sind inkonsistent/teilweise leer. Ziel: v2 mit drei Collections (notes/chunks/edges), neuem Chunker v2, Edge‑Builder v2 (explicit → rule → default_resolved → default), Policies (YAML), Observability (OTel), Snapshots. Kein Big‑Bang; v2 parallel zu v1.
|
||||
- Siehe angehängte Datei **mindnet_v2_implementation_playbook.md** (Pfad/Download im Chat). Alles daraus ist verbindlich.
|
||||
|
||||
Deine ersten Aufgaben (Step‑by‑Step):
|
||||
1) **Schemata anlegen (Step 1)**
|
||||
- Erstelle die Dateien:
|
||||
- schemas/note.schema.json
|
||||
- schemas/chunk.schema.json
|
||||
- schemas/edge.schema.json
|
||||
- schemas/default_edge.schema.json
|
||||
- Nutze die in der Playbook‑Datei vorgegebenen Strukturen als Basis.
|
||||
- Liefere zusätzlich ein einfaches `Makefile`‑Ziel `schema-validate` (jsonschema via Python), plus `requirements.txt`.
|
||||
- Output: komplette Dateien als Downloads + kurze Testanleitung.
|
||||
|
||||
2) **Qdrant v2‑Collections (Step 2)**
|
||||
- Erzeuge die Collections + Payload‑Indizes.
|
||||
- Liefere ein kleines Python‑Skript `tools/qdrant_bootstrap_v2.py`, das diese Anlage idempotent durchführt (inkl. Prüf‑Output).
|
||||
|
||||
3) **Chunker v2 (Step 3)**
|
||||
- Implementiere `chunking/block_chunker.py` mit den im Playbook genannten Parametern.
|
||||
- Teste gegen 3 Gold‑Notizen (liefere mini Test‑Harness in `tests/gold_notes/…`).
|
||||
|
||||
Arbeitsweise:
|
||||
- Jede Aufgabe einzeln, mit klaren Akzeptanzkriterien (aus Playbook) und Download‑Artefakten.
|
||||
- Keine parallelen Großumbauten.
|
||||
- Immer deterministische IDs, idempotente Upserts.
|
||||
|
||||
Dateien, die ich dir anfänglich bereitstelle:
|
||||
- **mindnet_v2_implementation_playbook.md** (diese Datei)
|
||||
- **import_markdown.py** (aktuelle Version aus dem Projekt)
|
||||
- **edges.py** (aktuelle Version aus dem Projekt)
|
||||
- **types.yaml** (aktuelle Version)
|
||||
|
||||
Sage mir jeweils, welche Datei du als Nächstes brauchst. Beginne jetzt mit Aufgabe 1 (Schemata).
|
||||
```
|
||||
@@ -0,0 +1,191 @@
|
||||
# mindnet v2.2 – Pipeline Playbook
|
||||
**Datei:** `docs/mindnet_pipeline_playbook_v2.2.md`
|
||||
**Stand:** 2025-12-07
|
||||
**Status:** **FINAL** (Konsolidiert aus WP02, WP03, WP04a)
|
||||
**Quellen:** `mindnet_v2_implementation_playbook.md`, `Handbuch.md`, `chunking_strategy.md`, `docs_mindnet_retriever.md`, `wp04_retriever_scoring.md`.
|
||||
|
||||
---
|
||||
|
||||
## 1. Zweck & Einordnung
|
||||
|
||||
Dieses Playbook ist das zentrale operative Handbuch für die **mindnet-Pipeline**. Es beschreibt, wie Daten vom Markdown-Vault in den Wissensgraphen (Qdrant) gelangen und wie der Retriever konfiguriert und betrieben wird.
|
||||
|
||||
**Zielgruppe:** Dev/Ops, Tech-Leads.
|
||||
**Scope:**
|
||||
* **Ist-Stand (WP01–WP04a):** Markdown-Import, Chunking, Edge-Erzeugung, Hybrider Retriever.
|
||||
* **Roadmap (Ausblick):** Self-Healing (WP06), Feedback-Loops (WP08).
|
||||
|
||||
---
|
||||
|
||||
## 2. Die Import-Pipeline (Runbook)
|
||||
|
||||
Der Import ist der kritischste Prozess. Er muss **deterministisch** und **idempotent** sein. Wir nutzen `scripts/import_markdown.py` als zentralen Entrypoint.
|
||||
|
||||
### 2.1 Der 12-Schritte-Prozess
|
||||
Gemäß WP03-Spezifikation läuft der Import intern wie folgt ab:
|
||||
|
||||
1. **Markdown lesen:** Rekursives Scannen des Vaults.
|
||||
2. **Frontmatter extrahieren:** Validierung von Pflichtfeldern (`id`, `type`, `title`).
|
||||
3. **Typauflösung:** Bestimmung des `type` via `types.yaml` (Prio: Frontmatter > Pfad > Default).
|
||||
4. **Note-Payload generieren:** Erstellen des JSON-Objekts für `mindnet_notes`.
|
||||
5. **Chunking anwenden:** Zerlegung des Textes basierend auf dem `chunk_profile` des Typs.
|
||||
6. **Inline-Kanten finden:** Parsing von `[[rel:...]]` im Fließtext.
|
||||
7. **Callout-Kanten finden:** Parsing von `> [!edge]` Blöcken.
|
||||
8. **Default-Edges erzeugen:** Anwendung der `edge_defaults` aus der Typ-Registry.
|
||||
9. **Strukturkanten erzeugen:** `belongs_to` (Chunk->Note), `next`/`prev` (Sequenz).
|
||||
10. **Chunks upserten:** Schreiben in Qdrant (`mindnet_chunks`).
|
||||
11. **Edges upserten:** Schreiben in Qdrant (`mindnet_edges`).
|
||||
12. **Diagnose:** Automatischer Check der Integrität nach dem Lauf.
|
||||
|
||||
### 2.2 Standard-Betrieb (Inkrementell)
|
||||
Für regelmäßige Updates (z.B. Cronjob). Erkennt Änderungen via Hash.
|
||||
|
||||
export QDRANT_URL="http://localhost:6333"
|
||||
export COLLECTION_PREFIX="mindnet"
|
||||
|
||||
# Import starten (Apply = Schreiben, Purge = Sauberer Upsert)
|
||||
python3 -m scripts.import_markdown \
|
||||
--vault ./vault \
|
||||
--prefix "$COLLECTION_PREFIX" \
|
||||
--apply \
|
||||
--purge-before-upsert \
|
||||
--sync-deletes
|
||||
|
||||
* `--apply`: Ohne dieses Flag läuft ein Dry-Run (nur Simulation).
|
||||
* `--purge-before-upsert`: Löscht vor dem Schreiben einer Note ihre alten Chunks/Edges. Essentiell, um "Geister-Chunks" zu vermeiden, wenn Text gekürzt wurde.
|
||||
* `--sync-deletes`: Entfernt Notizen aus Qdrant, die im Vault gelöscht wurden.
|
||||
|
||||
### 2.3 Full Rebuild (Clean Slate)
|
||||
Notwendig bei Änderungen an `types.yaml` (z.B. neue Chunk-Größen) oder Embedding-Modellen.
|
||||
|
||||
# 1. Qdrant Collections löschen und neu anlegen (Wipe inkl. Schema)
|
||||
python3 -m scripts.reset_qdrant --mode wipe --prefix "mindnet" --yes
|
||||
|
||||
# 2. Vollständiger Import aller Dateien
|
||||
python3 -m scripts.import_markdown \
|
||||
--vault ./vault \
|
||||
--prefix "mindnet" \
|
||||
--apply
|
||||
|
||||
### 2.4 Baseline-Builds & Hashing
|
||||
Um unnötige Updates zu vermeiden, nutzt der Importer Hashes.
|
||||
* **ENV `MINDNET_HASH_COMPARE`:** Steuert, was verglichen wird.
|
||||
* `Body`: Nur Textänderungen triggern Update (Standard).
|
||||
* `Full`: Auch Frontmatter-Änderungen (z.B. Tags) triggern Update.
|
||||
* **Flag `--baseline-modes`:** Berechnet Hashes für alle Modi vor, um spätere Wechsel der Strategie ohne Massen-Update zu ermöglichen.
|
||||
|
||||
---
|
||||
|
||||
## 3. Chunking & Payload-Aufbau
|
||||
|
||||
Das Chunking ist profilbasiert und typgesteuert.
|
||||
|
||||
### 3.1 Chunk-Profile
|
||||
In `types.yaml` definiert. Standard-Profile (in `chunk_config.py` implementiert):
|
||||
* `short`: Max 128 Tokens (z.B. für Logs, Chats).
|
||||
* `medium`: Max 256 Tokens (z.B. für Konzepte).
|
||||
* `long`: Max 512 Tokens (z.B. für Essays, Projekte).
|
||||
* `by_heading`: Trennt strikt an Überschriften.
|
||||
|
||||
### 3.2 Payload-Felder
|
||||
Jeder Chunk erhält zwei Text-Felder:
|
||||
* `text`: Der reine Inhalt des Chunks (ohne Overlap). Wird dem Nutzer angezeigt.
|
||||
* `window`: Der Inhalt plus Overlap zu Vorgänger/Nachfolger. Wird für das Embedding genutzt (besserer Kontext).
|
||||
|
||||
---
|
||||
|
||||
## 4. Edge-Erzeugung (Die V2-Logik)
|
||||
|
||||
In v2.2 entstehen Kanten nach strenger Priorität.
|
||||
|
||||
### 4.1 Prioritäten & Provenance
|
||||
Der Importer setzt `provenance`, `rule_id` und `confidence` automatisch:
|
||||
|
||||
| Priorität | Quelle | Syntax (Bsp.) | Rule ID | Confidence |
|
||||
| :--- | :--- | :--- | :--- | :--- |
|
||||
| **1** | Inline | `[[rel:depends_on X]]` | `inline:rel` | ~0.95 |
|
||||
| **2** | Callout | `> [!edge] related_to: [[X]]` | `callout:edge` | ~0.90 |
|
||||
| **3** | Wikilink | `[[X]]` | `explicit:wikilink` | 1.00 |
|
||||
| **4** | Default | *(via types.yaml)* | `edge_defaults:...` | ~0.70 |
|
||||
| **5** | Struktur | *(automatisch)* | `structure:...` | 1.00 |
|
||||
|
||||
### 4.2 Typ-Defaults
|
||||
Wenn in `types.yaml` für einen Typ `edge_defaults` definiert sind, werden diese **additiv** zu expliziten Links erzeugt.
|
||||
* *Beispiel:* Note Typ `project` verlinkt `[[Tool A]]`.
|
||||
* *Ergebnis:* Kante `references` (explizit) UND Kante `depends_on` (Default).
|
||||
|
||||
---
|
||||
|
||||
## 5. Retriever & Scoring
|
||||
|
||||
Der Retriever (`app/core/retriever.py`) kombiniert Signale zur Laufzeit.
|
||||
|
||||
### 5.1 Konfiguration (`retriever.yaml`)
|
||||
Diese Datei steuert das Ranking. Änderungen wirken sofort (API-Neustart).
|
||||
|
||||
scoring:
|
||||
semantic_weight: 1.0 # Vektor-Ähnlichkeit
|
||||
edge_weight: 0.7 # Graph-Dichte Bonus
|
||||
centrality_weight: 0.5 # Zentralitäts-Bonus
|
||||
|
||||
### 5.2 Scoring-Formel
|
||||
total_score =
|
||||
semantic_weight * semantic_score
|
||||
+ edge_weight * edge_bonus
|
||||
+ centrality_weight * centrality_bonus
|
||||
+ type_weight * retriever_weight
|
||||
|
||||
* `retriever_weight`: Kommt aus dem Note-Payload (via `types.yaml`).
|
||||
* `edge_bonus`: Summe der `confidence` aller relevanten Kanten im Subgraph.
|
||||
|
||||
### 5.3 Hybrider Modus
|
||||
Der Request muss `mode="hybrid"` und `expand.depth > 0` setzen, damit der Graph genutzt wird.
|
||||
* **Expand:** Lädt Nachbarn der Vektor-Treffer.
|
||||
* **Re-Rank:** Berechnet Boni auf diesem lokalen Graphen.
|
||||
|
||||
---
|
||||
|
||||
## 6. Quality Gates & Tests
|
||||
|
||||
Diese Tests garantieren die Stabilität der Pipeline.
|
||||
|
||||
### 6.1 Pflicht-Tests vor Commit
|
||||
|
||||
1. **Payload Dryrun (Schema-Check):**
|
||||
Simuliert Import, prüft JSON-Schema Konformität.
|
||||
|
||||
python3 -m scripts.payload_dryrun --vault ./test_vault
|
||||
|
||||
2. **Full Edge Check (Graph-Integrität):**
|
||||
Prüft Invarianten (z.B. `next` muss reziprok zu `prev` sein; keine verwaisten Kanten).
|
||||
|
||||
python3 -m scripts.edges_full_check
|
||||
|
||||
3. **Unit Tests:**
|
||||
|
||||
pytest tests/test_retriever_basic.py tests/test_retriever_edges.py
|
||||
|
||||
### 6.2 Smoke-Test (E2E)
|
||||
Prüft am laufenden System, ob Semantik und Graph funktionieren.
|
||||
|
||||
python scripts/test_retriever_smoke.py \
|
||||
--query "mindnet" \
|
||||
--mode hybrid \
|
||||
--expand-depth 1 \
|
||||
--top-k 5
|
||||
|
||||
### 6.3 Roundtrip-Test (Datenverlust-Check)
|
||||
Exportiert Qdrant zurück nach Markdown und vergleicht mit Original.
|
||||
|
||||
python3 -m scripts.export_markdown --out ./_export
|
||||
python3 tests/compare_vaults.py --src ./vault --dst ./_export --focus body
|
||||
|
||||
---
|
||||
|
||||
## 7. Ausblick & Roadmap
|
||||
|
||||
### 7.1 Self-Healing (WP06)
|
||||
Geplant: Automatisierte Jobs, die `unresolved` Referenzen periodisch prüfen und heilen, wenn die Ziel-Note erstellt wurde. Aktuell manuell via `scripts/resolve_unresolved_references.py`.
|
||||
|
||||
### 7.2 Feedback-Logging (WP04c/WP08)
|
||||
Geplant: Speicherung von User-Feedback zu Suchergebnissen, um `retriever.yaml` Gewichte mittelfristig automatisch zu justieren ("Self-Tuning").
|
||||
@@ -0,0 +1,114 @@
|
||||
# Mindnet v2.2 – User Guide
|
||||
**Datei:** `docs/mindnet_user_guide_v2.2.md`
|
||||
**Stand:** 2025-12-07
|
||||
**Status:** **FINAL** (Konsolidiert WP01–WP04a)
|
||||
**Quellen:** `knowledge_design.md`, `wp04_retriever_scoring.md`, `Programmplan_V2.2.md`, `Handbuch.md`.
|
||||
|
||||
> **Willkommen bei Mindnet.**
|
||||
> Dies ist dein persönliches Wissensnetzwerk. Im Gegensatz zu einer normalen Suche (wie Google Drive oder Spotlight), die nur Texte findet, "denkt" Mindnet mit. Dieser Guide zeigt dir, wie du das System nutzt.
|
||||
|
||||
---
|
||||
|
||||
## 1. Was Mindnet für dich tut
|
||||
|
||||
Mindnet ist ein **assoziatives Gedächtnis**. Es verbindet deine Notizen zu einem Graphen.
|
||||
|
||||
### 1.1 Mehr als nur Stichworte
|
||||
Wenn du nach "Projekt Alpha" suchst, findet Mindnet nicht nur das Dokument mit diesem Titel. Es findet auch:
|
||||
* **Abhängigkeiten:** "Technologie X wird benötigt" (weil sie verlinkt ist).
|
||||
* **Entscheidungen:** "Warum nutzen wir X?" (weil eine Decision-Note existiert).
|
||||
* **Ähnliches:** "Projekt Beta war ähnlich" (weil es strukturell verwandt ist).
|
||||
|
||||
### 1.2 Dein KI-Zwilling (Vision)
|
||||
Mindnet speichert nicht nur Fakten, sondern deine **Perspektive**. Durch die Gewichtung von Typen (z.B. "Prinzipien sind wichtiger als Quellen") lernt das System, Antworten so zu priorisieren, wie du es tun würdest.
|
||||
|
||||
---
|
||||
|
||||
## 2. Gute Fragen an Mindnet stellen
|
||||
|
||||
Mindnet nutzt eine **Hybride Suche**. Das heißt, es schaut auf deine Worte (Semantik) und auf das Netzwerk (Graph).
|
||||
|
||||
### 2.1 Frage-Muster
|
||||
Um gute Antworten zu erhalten, formuliere deine Anfragen präzise:
|
||||
|
||||
* **Faktensuche:** "Wie installiere ich Qdrant?"
|
||||
* *Mindnet sucht:* Chunks mit technischer Anleitung. Hier gewinnt oft die reine Text-Ähnlichkeit.
|
||||
* **Zusammenhänge:** "Womit hängt Projekt Alpha zusammen?"
|
||||
* *Mindnet sucht:* Den Projekt-Knoten und folgt den Kanten (`depends_on`, `related_to`). Hier gewinnt der Graph.
|
||||
* **Entscheidungen:** "Warum nutzen wir Vektordatenbanken?"
|
||||
* *Mindnet sucht:* Notizen vom Typ `decision` oder `principle`. Hier gewinnt das Typ-Gewicht (`retriever_weight`).
|
||||
|
||||
### 2.2 Tipps für bessere Ergebnisse
|
||||
1. **Nutze deine Sprache:** Verwende die Begriffe, die du auch in deinen Notizen benutzt hast.
|
||||
2. **Kontext geben:** Statt nur "Qdrant" zu tippen (was alles finden würde), frage "Qdrant Setup für Mindnet", um den Kontext einzuschränken.
|
||||
|
||||
---
|
||||
|
||||
## 3. Ergebnisse interpretieren
|
||||
|
||||
Wenn Mindnet antwortet, siehst du oft mehr als nur Text. Du siehst eine **Begründung**.
|
||||
|
||||
### 3.1 Der Score (Warum ist das hier oben?)
|
||||
Mindnet berechnet für jeden Treffer einen `total_score`. Dieser ist keine Magie, sondern Mathematik:
|
||||
|
||||
Score = (Text-Treffer) + (Wichtigkeit) + (Vernetzung)
|
||||
|
||||
* **Text-Treffer (Semantik):** Wie gut passen die Worte?
|
||||
* **Wichtigkeit (Typ):** Ist die Quelle eine zentrale `decision` (hoch gewichtet) oder nur eine `source` (niedrig)? Ein Treffer in einer Entscheidung schlägt oft einen Treffer in einem Buch-Exzerpt.
|
||||
* **Vernetzung (Edge Bonus):** Wird dieser Inhalt oft von anderen wichtigen Projekten referenziert? Ein "zentraler" Knoten wird bevorzugt.
|
||||
|
||||
### 3.2 Pfade (Der Kontext)
|
||||
Oft zeigt Mindnet im Hintergrund (Explainability):
|
||||
*"Gefunden via: Projekt Alpha -> depends_on -> Vektordatenbank"*
|
||||
|
||||
Das bedeutet: Dieser Treffer ist relevant, weil er eine **Abhängigkeit** deines aktuellen Projekts ist, auch wenn deine Suchworte vielleicht nicht exakt passen. Das ist der Moment, in dem Mindnet "mitdenkt".
|
||||
|
||||
---
|
||||
|
||||
## 4. Neues Wissen beisteuern (Authoring)
|
||||
|
||||
Mindnet lebt von deinem Input. Du musst kein Techniker sein, um gutes Wissen zu designen. Du schreibst einfach Markdown.
|
||||
|
||||
### 4.1 Die Goldene Regel: "Verlinke semantisch"
|
||||
Statt einfach nur `[[Link]]` zu schreiben, versuche zu sagen, *wie* es zusammenhängt. Das hilft dem Retriever später, den "Pfad" zu finden.
|
||||
|
||||
* Hängt es davon ab? -> `[[rel:depends_on Ziel]]`
|
||||
* Ist es ähnlich? -> `[[rel:similar_to Ziel]]`
|
||||
* Ist es eine Folge davon? -> `[[rel:caused_by Ziel]]`
|
||||
|
||||
### 4.2 Struktur hilft dem Chunker
|
||||
Mindnet zerlegt deinen Text in Häppchen ("Chunks"). Hilf dabei:
|
||||
* Verwende **Überschriften** (##), um Themen zu trennen.
|
||||
* Schreibe **Absätze**, die einen Gedanken fassen.
|
||||
* Vermeide endlose Textwüsten.
|
||||
|
||||
### 4.3 Typen nutzen
|
||||
Setze im Frontmatter deiner Notizen den richtigen Typ.
|
||||
* `type: project` -> Mindnet weiß: Das hat Aufgaben und Abhängigkeiten.
|
||||
* `type: decision` -> Mindnet weiß: Das ist wichtig und begründet Dinge.
|
||||
* `type: journal` -> Mindnet weiß: Das ist zeitgebunden und weniger wichtig für Grundsatzfragen.
|
||||
|
||||
**Beispiel:**
|
||||
---
|
||||
title: Entscheidung für Python
|
||||
type: decision
|
||||
---
|
||||
Wir nutzen Python, weil es [[rel:solves AI Library Requirements]].
|
||||
|
||||
---
|
||||
|
||||
## 5. Feedback & Lernen
|
||||
|
||||
Mindnet wird schlauer, wenn du es pflegst.
|
||||
|
||||
### 5.1 Ergebnis fehlt?
|
||||
* Existiert die Notiz im Vault?
|
||||
* Ist sie isoliert (keine Links)? Isolierte Notizen haben keinen Graph-Bonus und werden schlechter gefunden. **Verlinke sie!**
|
||||
* Ist der Typ `source`? Dann ist das Gewicht niedrig (0.5). Wenn es deine eigene Meinung ist, ändere den Typ auf `concept` oder `experience`.
|
||||
|
||||
### 5.2 Falsches Ergebnis?
|
||||
* Ist die Notiz veraltet? Setze `status: archived` (wenn du Filter konfiguriert hast) oder lösche sie.
|
||||
* Wird ein Begriff falsch verstanden? Füge Synonyme oder eine Definition hinzu.
|
||||
|
||||
### 5.3 Ausblick (Roadmap)
|
||||
In Zukunft (WP08) wirst du Ergebnisse direkt mit "Daumen hoch/runter" bewerten können. Mindnet passt dann seine Gewichte (`retriever.yaml`) automatisch an, um deine Präferenzen zu lernen.
|
||||
@@ -0,0 +1,260 @@
|
||||
# WP-04 / Step 4a – Retriever Scoring & Konfiguration
|
||||
|
||||
Dieses Dokument beschreibt den aktuellen Stand (2025-11-30) der Scoring-Logik des mindnet-Retrievers in WP-04 / Step 4a.
|
||||
Es dient als Referenz für:
|
||||
|
||||
- Nachvollziehbarkeit des Rankings
|
||||
- Konfiguration und Feintuning (ohne Codeänderungen)
|
||||
- spätere Automatisierung / „Selbstjustierung“ des Retrievers
|
||||
|
||||
---
|
||||
|
||||
## 1. Überblick
|
||||
|
||||
Der Retriever kombiniert drei Signalquellen zu einem einheitlichen Score:
|
||||
|
||||
1. **Semantik** – Vektorsuche über `mindnet_chunks`
|
||||
2. **Typ-Gewichtung** – `retriever_weight` pro Note/Chunk (aus `types.yaml`)
|
||||
3. **Graph-Boni** – `edge_bonus` und `centrality_bonus` aus dem Subgraph (Edges)
|
||||
|
||||
Die Berechnung erfolgt in `app/core/retriever.py` und ist über `config/retriever.yaml` konfigurierbar.
|
||||
Die API stellt die Funktionalität über den `/query`-Endpoint (FastAPI) bereit.
|
||||
|
||||
---
|
||||
|
||||
## 2. Datenbasis (Kurzüberblick)
|
||||
|
||||
### 2.1 Notes (`<prefix>_notes`)
|
||||
|
||||
Relevante Felder:
|
||||
|
||||
- note_id: keyword
|
||||
- title: text
|
||||
- type: keyword
|
||||
- retriever_weight: float
|
||||
- chunk_profile: keyword
|
||||
- edge_defaults: keyword[]
|
||||
- tags: keyword[]
|
||||
- path: text
|
||||
- fulltext: text
|
||||
|
||||
### 2.2 Chunks (`<prefix>_chunks`)
|
||||
|
||||
Relevante Felder:
|
||||
|
||||
- note_id: keyword
|
||||
- chunk_id: keyword
|
||||
- text: text
|
||||
- window: text
|
||||
- retriever_weight: float
|
||||
- type: keyword
|
||||
- path: text
|
||||
- section: text
|
||||
- neighbors_prev / neighbors_next
|
||||
- chunk_profile: keyword
|
||||
|
||||
### 2.3 Edges (`<prefix>_edges`)
|
||||
|
||||
Relevante Felder:
|
||||
|
||||
- kind: keyword
|
||||
- source_id: keyword
|
||||
- target_id: keyword
|
||||
- note_id: keyword
|
||||
- confidence: float
|
||||
- rule_id, provenance, edge_id, scope, relation, ref_text
|
||||
|
||||
Edges stammen stets aus der Import-/Edge-Pipeline (WP-03) und bilden die graphbasierte Grundlage des Retrievers.
|
||||
|
||||
---
|
||||
|
||||
## 3. Scoring-Formel (Retriever)
|
||||
|
||||
### 3.1 Eingangsgrößen
|
||||
|
||||
Für jeden Treffer gelten:
|
||||
|
||||
semantic_score – aus der Vektorsuche (Qdrant)
|
||||
retriever_weight – typabhängiges Gewicht (aus types.yaml)
|
||||
edge_bonus – aggregierter Graph-Bonus
|
||||
centrality_bonus – Zentralitätsbonus aus dem Subgraph
|
||||
|
||||
### 3.2 Globale Gewichte (aus retriever.yaml)
|
||||
|
||||
Datei: `config/retriever.yaml`
|
||||
Beispiel:
|
||||
|
||||
version: 1.0
|
||||
|
||||
scoring:
|
||||
semantic_weight: 1.0 # W_sem
|
||||
edge_weight: 0.7 # W_edge
|
||||
centrality_weight: 0.5 # W_cent
|
||||
|
||||
Falls diese Datei fehlt, greifen ENV-basierte Defaults (`RETRIEVER_W_*`).
|
||||
|
||||
### 3.3 Formel
|
||||
|
||||
total_score =
|
||||
W_sem * semantic_score * max(retriever_weight, 0.0)
|
||||
+ W_edge * edge_bonus
|
||||
+ W_cent * centrality_bonus
|
||||
|
||||
Bemerkungen:
|
||||
|
||||
- Negative `retriever_weight` werden abgefangen.
|
||||
- Die rohen `edge_bonus` und `centrality_bonus` werden im API-/Smoke-Output angezeigt.
|
||||
- Die Gewichte aus `retriever.yaml` beeinflussen **nur den total_score**, nicht die Anzeige.
|
||||
|
||||
---
|
||||
|
||||
## 4. Konfigurationsdateien
|
||||
|
||||
### 4.1 Typen-Registry (`types.yaml`)
|
||||
|
||||
Beispielauszug:
|
||||
|
||||
defaults:
|
||||
retriever_weight: 1.0
|
||||
chunk_profile: default
|
||||
|
||||
types:
|
||||
types:
|
||||
concept:
|
||||
retriever_weight: 0.60
|
||||
edge_defaults: ["references", "related_to"]
|
||||
|
||||
project:
|
||||
retriever_weight: 0.97
|
||||
edge_defaults: ["references", "depends_on"]
|
||||
|
||||
journal:
|
||||
retriever_weight: 0.80
|
||||
edge_defaults: ["references", "related_to"]
|
||||
|
||||
source:
|
||||
retriever_weight: 0.50
|
||||
edge_defaults: []
|
||||
|
||||
Diese Werte bestimmen, wie wichtig eine Note/Chunk **prinzipiell** für den Retriever ist.
|
||||
|
||||
### 4.2 Retriever-Konfiguration (`retriever.yaml`)
|
||||
|
||||
Beispiel:
|
||||
|
||||
version: 1.0
|
||||
|
||||
scoring:
|
||||
semantic_weight: 1.0
|
||||
edge_weight: 0.7
|
||||
centrality_weight: 0.5
|
||||
|
||||
# Blueprint für spätere Edge-Feinsteuerung:
|
||||
edge_types:
|
||||
references: 0.20
|
||||
depends_on: 0.18
|
||||
related_to: 0.15
|
||||
similar_to: 0.12
|
||||
belongs_to: 0.10
|
||||
next: 0.06
|
||||
prev: 0.06
|
||||
|
||||
Die Datei befindet sich standardmäßig unter `config/retriever.yaml`.
|
||||
Override per ENV:
|
||||
|
||||
MINDNET_RETRIEVER_CONFIG=/pfad/zur/datei.yaml
|
||||
|
||||
### 4.3 Aktive Gewichte auslesen
|
||||
|
||||
python - << 'PY'
|
||||
from app.core.retriever import _get_scoring_weights
|
||||
print(_get_scoring_weights())
|
||||
PY
|
||||
|
||||
Beispielausgabe:
|
||||
|
||||
(1.0, 0.7, 0.5)
|
||||
|
||||
---
|
||||
|
||||
## 5. Betriebsmodi
|
||||
|
||||
### 5.1 Semantikmodus (semantic)
|
||||
|
||||
Ablauf:
|
||||
|
||||
- Query-Vektor über embed_text()
|
||||
- Qdrant-Suche über mindnet_chunks
|
||||
- kein Subgraph → edge_bonus = centrality = 0
|
||||
- total_score = W_sem * semantic_score * retriever_weight
|
||||
|
||||
### 5.2 Hybridmodus (hybrid)
|
||||
|
||||
Ablauf:
|
||||
|
||||
- Semantik wie oben
|
||||
- falls expand.depth > 0:
|
||||
- Subgraph-Berechnung über ga.expand()
|
||||
- Ermittlung von edge_bonus und centrality_bonus
|
||||
- total_score = kombinierte Formel aus Semantik + Graph
|
||||
|
||||
---
|
||||
|
||||
## 6. Tests
|
||||
|
||||
### Ausführung (Unit & Integration)
|
||||
|
||||
pytest tests/test_retriever_basic.py \
|
||||
tests/test_retriever_weight.py \
|
||||
tests/test_retriever_edges.py
|
||||
|
||||
pytest tests/test_query_unit.py \
|
||||
tests/test_query_text_embed_unit.py
|
||||
|
||||
Tests decken ab:
|
||||
|
||||
- Semantiktreffer
|
||||
- Typ-Gewichte
|
||||
- Edge-Boni
|
||||
- Hybridmodus / FastAPI-Endpunkt
|
||||
|
||||
### Smoke-Test
|
||||
|
||||
python tests/test_retriever_smoke.py \
|
||||
--url "http://127.0.0.1:8001/query" \
|
||||
--query "embeddings" \
|
||||
--mode hybrid \
|
||||
--expand-depth 1 \
|
||||
--top-k 5
|
||||
|
||||
Beispielausgabe:
|
||||
|
||||
[1] note_id=...retriever-design
|
||||
total=1.5054 (semantic=0.4038 edge=1.5910 centrality=0.0000)
|
||||
|
||||
Interpretation:
|
||||
|
||||
- `semantic`, `edge`, `centrality` = Rohwerte
|
||||
- `total` = skaliert mit `retriever.yaml`
|
||||
|
||||
---
|
||||
|
||||
## 7. Betriebs-Hinweise
|
||||
|
||||
### 7.1 Änderung der Gewichte
|
||||
|
||||
- retriever.yaml anpassen
|
||||
- FastAPI/uvicorn neu starten
|
||||
- Smoke-Test prüfen
|
||||
|
||||
### 7.2 Zielbild WP-04
|
||||
|
||||
Der jetzige Stand ermöglicht:
|
||||
|
||||
- transparente, konfigurierbare Score-Berechnung
|
||||
- Kombination aus Semantik, Typwissen und Graphstruktur
|
||||
- spätere agentenbasierte Selbstjustierung
|
||||
- alternative Profile (z. B. „mehr Semantik“, „mehr Graph“)
|
||||
|
||||
Dieses Dokument dient als stabile Referenz für Entwickler, Tests und spätere Automatisierungsschritte.
|
||||
|
||||
+35
-10
@@ -1,10 +1,35 @@
|
||||
fastapi
|
||||
uvicorn[standard]
|
||||
qdrant-client
|
||||
pydantic>=2
|
||||
sentence-transformers
|
||||
torch
|
||||
python-frontmatter
|
||||
tqdm
|
||||
requests
|
||||
numpy
|
||||
# --- API Server ---
|
||||
fastapi>=0.116.1
|
||||
uvicorn[standard]>=0.35.0
|
||||
# httpx wird oft für Tests benötigt
|
||||
httpx>=0.28.1
|
||||
|
||||
# --- Datenbank ---
|
||||
qdrant-client>=1.15.1
|
||||
|
||||
# --- Datenmodelle & Validierung ---
|
||||
pydantic>=2.11.7
|
||||
numpy>=2.3.2
|
||||
|
||||
# --- Markdown & Parsing (Hier fehlten Pakete!) ---
|
||||
python-frontmatter>=1.1.0
|
||||
# WICHTIG: Das fehlte und verursachte den Fehler
|
||||
markdown-it-py>=3.0.0
|
||||
# WICHTIG: Für types.yaml und retriever.yaml
|
||||
PyYAML>=6.0.2
|
||||
python-slugify>=8.0.4
|
||||
|
||||
# --- KI & Embeddings ---
|
||||
sentence-transformers>=5.1.0
|
||||
# Torch wird meist durch sentence-transformers geholt,
|
||||
# aber wir listen es explizit für Stabilität
|
||||
torch>=2.0.0
|
||||
|
||||
# --- Utilities ---
|
||||
# WICHTIG: Damit .env Dateien gelesen werden
|
||||
python-dotenv>=1.1.1
|
||||
requests>=2.32.5
|
||||
tqdm>=4.67.1
|
||||
|
||||
# --- Testing ---
|
||||
pytest>=8.4.2
|
||||
@@ -0,0 +1,116 @@
|
||||
#!/usr/bin/env python3
|
||||
# -*- coding: utf-8 -*-
|
||||
"""
|
||||
tests/assert_payload_schema.py (compat v1.1)
|
||||
|
||||
Prüft, ob die erwarteten Payload-Indizes (payload_schema) auf den drei Collections vorhanden sind.
|
||||
Kompatibel mit älteren qdrant-client Versionen (ohne with_payload_schema im Wrapper).
|
||||
|
||||
Collections & Pflichtfelder:
|
||||
- mindnet_notes : note_id, type, title, updated, tags
|
||||
- mindnet_chunks : note_id, chunk_id, index, type, tags
|
||||
- mindnet_edges : note_id, kind, scope, source_id, target_id, chunk_id
|
||||
|
||||
Nutzung:
|
||||
python3 -m tests.assert_payload_schema
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import os
|
||||
import sys
|
||||
import urllib.error
|
||||
import urllib.request
|
||||
from typing import Any, Dict, List
|
||||
|
||||
from app.core.qdrant import QdrantConfig, get_client, collection_names
|
||||
|
||||
REQUIRED = {
|
||||
"notes": ["note_id", "type", "title", "updated", "tags"],
|
||||
"chunks": ["note_id", "chunk_id", "index", "type", "tags"],
|
||||
"edges": ["note_id", "kind", "scope", "source_id", "target_id", "chunk_id"],
|
||||
}
|
||||
|
||||
def _safe_model_dump(obj: Any) -> Dict[str, Any]:
|
||||
if hasattr(obj, "model_dump"):
|
||||
return obj.model_dump()
|
||||
if hasattr(obj, "dict"):
|
||||
return obj.dict()
|
||||
return obj if isinstance(obj, dict) else {}
|
||||
|
||||
def _cfg_base_url(cfg) -> str:
|
||||
if getattr(cfg, "url", None):
|
||||
return cfg.url.rstrip('/')
|
||||
host = getattr(cfg, "host", None) or os.getenv("QDRANT_HOST") or "127.0.0.1"
|
||||
port = getattr(cfg, "port", None) or os.getenv("QDRANT_PORT") or "6333"
|
||||
return f"http://{host}:{port}"
|
||||
|
||||
def _http_get(url: str, api_key: str | None) -> Dict[str, Any]:
|
||||
req = urllib.request.Request(url, method="GET")
|
||||
if api_key:
|
||||
req.add_header("api-key", api_key)
|
||||
req.add_header("Authorization", f"Bearer {api_key}")
|
||||
req.add_header("Accept", "application/json")
|
||||
with urllib.request.urlopen(req, timeout=30) as resp:
|
||||
data = resp.read()
|
||||
return json.loads(data.decode("utf-8"))
|
||||
|
||||
def get_collection_payload_schema(client, cfg, name: str) -> Dict[str, Any] | None:
|
||||
# 1) OpenAPI client mit Flag
|
||||
try:
|
||||
oc = getattr(client, "openapi_client", None)
|
||||
if oc is not None and hasattr(oc, "collections_api"):
|
||||
api = oc.collections_api
|
||||
info = api.get_collection(collection_name=name, with_payload_schema=True)
|
||||
d = _safe_model_dump(info)
|
||||
return (d.get("result") or {}).get("payload_schema")
|
||||
except Exception:
|
||||
pass
|
||||
# 2) Wrapper (manche Versionen liefern Schema auch ohne Flag)
|
||||
try:
|
||||
info = client.get_collection(collection_name=name)
|
||||
d = _safe_model_dump(info)
|
||||
ps = (d.get("result") or {}).get("payload_schema")
|
||||
if ps is not None:
|
||||
return ps
|
||||
except Exception:
|
||||
pass
|
||||
# 3) Direkter HTTP-Call
|
||||
try:
|
||||
base = _cfg_base_url(cfg)
|
||||
url = f"{base}/collections/{name}?with_payload_schema=true"
|
||||
raw = _http_get(url, getattr(cfg, "api_key", None))
|
||||
return (raw.get("result") or {}).get("payload_schema")
|
||||
except Exception:
|
||||
return None
|
||||
|
||||
def ensure_fields(schema: Dict[str, dict] | None, required: List[str]) -> Dict[str, bool]:
|
||||
ks = set((schema or {}).keys())
|
||||
return {k: (k in ks) for k in required}
|
||||
|
||||
def main():
|
||||
cfg = QdrantConfig.from_env()
|
||||
client = get_client(cfg)
|
||||
notes, chunks, edges = collection_names(cfg.prefix)
|
||||
names = {"notes": notes, "chunks": chunks, "edges": edges}
|
||||
|
||||
report = {}
|
||||
ok_all = True
|
||||
for kind, col in names.items():
|
||||
sch = get_collection_payload_schema(client, cfg, col)
|
||||
checks = ensure_fields(sch, REQUIRED[kind])
|
||||
ok = all(checks.values())
|
||||
ok_all = ok_all and ok
|
||||
report[kind] = {
|
||||
"collection": col,
|
||||
"ok": ok,
|
||||
"missing": [k for k,v in checks.items() if not v],
|
||||
"present": [k for k,v in checks.items() if v],
|
||||
}
|
||||
|
||||
print(json.dumps({"prefix": cfg.prefix, "ok": ok_all, "report": report}, ensure_ascii=False, indent=2))
|
||||
if not ok_all:
|
||||
sys.exit(1)
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -0,0 +1,102 @@
|
||||
#!/usr/bin/env python3
|
||||
# -*- coding: utf-8 -*-
|
||||
"""
|
||||
scripts/diag_payload_indexes.py (compat v1.2)
|
||||
|
||||
Zeigt payload_schema je Collection; kompatibel mit älteren qdrant-client Versionen.
|
||||
Strategie (in dieser Reihenfolge):
|
||||
1) openapi_client.collections_api.get_collection(..., with_payload_schema=True)
|
||||
2) client.get_collection(...)
|
||||
3) HTTP GET /collections/{name}?with_payload_schema=true
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import os
|
||||
import urllib.request
|
||||
import urllib.error
|
||||
from typing import Any, Dict
|
||||
|
||||
from app.core.qdrant import QdrantConfig, get_client, collection_names
|
||||
|
||||
|
||||
def _safe_model_dump(obj: Any) -> Dict[str, Any]:
|
||||
if hasattr(obj, "model_dump"):
|
||||
return obj.model_dump()
|
||||
if hasattr(obj, "dict"):
|
||||
return obj.dict()
|
||||
return obj if isinstance(obj, dict) else {}
|
||||
|
||||
|
||||
def _cfg_base_url(cfg) -> str:
|
||||
if getattr(cfg, "url", None):
|
||||
return cfg.url.rstrip("/")
|
||||
host = getattr(cfg, "host", None) or os.getenv("QDRANT_HOST") or "127.0.0.1"
|
||||
port = getattr(cfg, "port", None) or os.getenv("QDRANT_PORT") or "6333"
|
||||
return f"http://{host}:{port}"
|
||||
|
||||
|
||||
def _http_get(url: str, api_key: str | None) -> Dict[str, Any]:
|
||||
req = urllib.request.Request(url, method="GET")
|
||||
if api_key:
|
||||
req.add_header("api-key", api_key)
|
||||
req.add_header("Authorization", f"Bearer {api_key}")
|
||||
req.add_header("Accept", "application/json")
|
||||
with urllib.request.urlopen(req, timeout=30) as resp:
|
||||
data = resp.read()
|
||||
return json.loads(data.decode("utf-8"))
|
||||
|
||||
|
||||
def get_collection_payload_schema(client, cfg, name: str) -> Dict[str, Any] | None:
|
||||
# Strategy 1: OpenAPI client with flag
|
||||
try:
|
||||
oc = getattr(client, "openapi_client", None)
|
||||
if oc is not None and hasattr(oc, "collections_api"):
|
||||
api = oc.collections_api
|
||||
info = api.get_collection(collection_name=name, with_payload_schema=True)
|
||||
d = _safe_model_dump(info)
|
||||
return (d.get("result") or {}).get("payload_schema")
|
||||
except Exception:
|
||||
pass
|
||||
# Strategy 2: wrapper (manche Versionen liefern Schema auch ohne Flag)
|
||||
try:
|
||||
info = client.get_collection(collection_name=name) # kein with_payload_schema kwarg!
|
||||
d = _safe_model_dump(info)
|
||||
ps = (d.get("result") or {}).get("payload_schema")
|
||||
if ps is not None:
|
||||
return ps
|
||||
except Exception:
|
||||
pass
|
||||
# Strategy 3: direkter HTTP Call
|
||||
try:
|
||||
base = _cfg_base_url(cfg)
|
||||
url = f"{base}/collections/{name}?with_payload_schema=true"
|
||||
raw = _http_get(url, getattr(cfg, "api_key", None))
|
||||
return (raw.get("result") or {}).get("payload_schema")
|
||||
except Exception:
|
||||
return None
|
||||
|
||||
|
||||
def main():
|
||||
cfg = QdrantConfig.from_env()
|
||||
client = get_client(cfg)
|
||||
notes, chunks, edges = collection_names(cfg.prefix)
|
||||
cols = [notes, chunks, edges]
|
||||
out = []
|
||||
for name in cols:
|
||||
ps = get_collection_payload_schema(client, cfg, name)
|
||||
# Vektorinfos (optional)
|
||||
try:
|
||||
info = client.get_collection(collection_name=name)
|
||||
d = _safe_model_dump(info)
|
||||
vectors = (d.get("result") or {}).get("vectors")
|
||||
if isinstance(vectors, dict) and "config" in vectors:
|
||||
vectors = vectors["config"]
|
||||
except Exception:
|
||||
vectors = None
|
||||
out.append({"name": name, "vectors": vectors, "payload_schema": ps})
|
||||
print(json.dumps({"prefix": cfg.prefix, "collections": out}, ensure_ascii=False, indent=2))
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -0,0 +1,107 @@
|
||||
#!/usr/bin/env python3
|
||||
# scripts/edges_dryrun.py
|
||||
# -*- coding: utf-8 -*-
|
||||
"""
|
||||
Dry-Run: Erzeuge Edges aus einem Vault **ohne** Qdrant-Upsert.
|
||||
- Liest Markdown mit YAML-Frontmatter
|
||||
- Chunking: einfacher Absatz-Chunker (Index + text)
|
||||
- Kanten: nutzt app.core.edges.build_edges_for_note()
|
||||
- Ausgabe: JSON pro Note mit Edge-Counts und 3 Beispiel-Payloads
|
||||
|
||||
Aufruf:
|
||||
python3 -m scripts.edges_dryrun --vault ./vault
|
||||
|
||||
Optional:
|
||||
--include-note-scope-refs # auch Frontmatter-Links (links[].target_id) als Note-Scope-Referenzen
|
||||
|
||||
Voraussetzungen:
|
||||
- app/core/parser.py (read_markdown, normalize_frontmatter)
|
||||
- app/core/edges.py (dieses Modul forwardet zur v2-Implementierung)
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import argparse
|
||||
import json
|
||||
import os
|
||||
import re
|
||||
from pathlib import Path
|
||||
from typing import Dict, List, Optional
|
||||
|
||||
from app.core.parser import read_markdown, normalize_frontmatter
|
||||
from app.core.edges import build_edges_for_note
|
||||
|
||||
def _iter_markdown(vault: str):
|
||||
for p in Path(vault).rglob("*.md"):
|
||||
if p.name.startswith("."): # ignore hidden
|
||||
continue
|
||||
yield p
|
||||
|
||||
def _simple_chunker(body: str, note_id: str, note_type: str) -> List[Dict]:
|
||||
# Absatzbasiert, minimal ausreichend für Edges (window/text, chunk_id, ord, note_id, type)
|
||||
paras = [s.strip() for s in re.split(r"\n{2,}", body or "") if s.strip()]
|
||||
chunks = []
|
||||
for i, t in enumerate(paras):
|
||||
chunks.append({
|
||||
"chunk_id": f"{note_id}#c{i:04d}",
|
||||
"ord": i,
|
||||
"text": t,
|
||||
"note_id": note_id,
|
||||
"type": note_type,
|
||||
})
|
||||
return chunks
|
||||
|
||||
def _fm_note_refs(fm: Dict) -> List[str]:
|
||||
out = []
|
||||
links = fm.get("links") or []
|
||||
if isinstance(links, list):
|
||||
for e in links:
|
||||
if isinstance(e, dict):
|
||||
tid = e.get("target_id")
|
||||
if isinstance(tid, str) and tid.strip():
|
||||
out.append(tid.strip())
|
||||
return out
|
||||
|
||||
def main():
|
||||
ap = argparse.ArgumentParser()
|
||||
ap.add_argument("--vault", required=True)
|
||||
ap.add_argument("--include-note-scope-refs", action="store_true")
|
||||
args = ap.parse_args()
|
||||
|
||||
vault = args.vault
|
||||
include_note_scope = args.include_note_scope_refs
|
||||
|
||||
report = []
|
||||
for p in _iter_markdown(vault):
|
||||
parsed = read_markdown(str(p))
|
||||
if not parsed:
|
||||
continue
|
||||
fm = normalize_frontmatter(parsed.frontmatter or {})
|
||||
note_id = fm.get("id") or p.stem
|
||||
note_type = (fm.get("type") or "concept").lower()
|
||||
chunks = _simple_chunker(parsed.body, note_id, note_type)
|
||||
note_refs = _fm_note_refs(fm)
|
||||
|
||||
edges = build_edges_for_note(
|
||||
note_id=note_id,
|
||||
chunk_payloads=chunks,
|
||||
note_level_refs=note_refs,
|
||||
include_note_scope_refs=include_note_scope,
|
||||
)
|
||||
kinds = {}
|
||||
for e in edges:
|
||||
key = (e.get("relation") or e.get("kind") or "edge")
|
||||
kinds[key] = kinds.get(key, 0) + 1
|
||||
report.append({
|
||||
"path": str(p),
|
||||
"note_id": note_id,
|
||||
"type": note_type,
|
||||
"chunks": len(chunks),
|
||||
"edges_total": len(edges),
|
||||
"edges_by_kind": kinds,
|
||||
"samples": edges[:3],
|
||||
})
|
||||
|
||||
print(json.dumps(report, ensure_ascii=False, indent=2))
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -0,0 +1,163 @@
|
||||
#!/usr/bin/env python3
|
||||
# -*- coding: utf-8 -*-
|
||||
"""
|
||||
scripts/edges_full_check.py
|
||||
Zählt und validiert Kanten in Qdrant. Erkennt folgende Rule-Gruppen:
|
||||
- explicit_total: rule_id startswith "explicit:" (z.B. explicit:wikilink, explicit:note_scope)
|
||||
- callout_total: rule_id == "callout:edge"
|
||||
- inline_total: rule_id startswith "inline:" (z.B. inline:rel)
|
||||
- defaults_total: rule_id startswith "edge_defaults:"
|
||||
- structure: rule_id in {"structure:belongs_to","structure:order"}
|
||||
|
||||
Gibt zusätzlich:
|
||||
- edges_by_kind (aggregiert)
|
||||
- notes/chunks/edges Anzahlen
|
||||
- multi_callout_detected: True, falls ein Chunk mehrere Callout-Ziele der gleichen Relation enthält
|
||||
- per_note_checks: belongs_to == chunks, next == prev == (chunks-1)
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
from collections import Counter, defaultdict
|
||||
from typing import Dict, Any, List, Tuple
|
||||
|
||||
from app.core.qdrant import QdrantConfig, get_client
|
||||
from qdrant_client.http import models as rest
|
||||
|
||||
|
||||
def _count_collection_points(client, name: str) -> int:
|
||||
try:
|
||||
res = client.count(collection_name=name, exact=True)
|
||||
return res.count or 0
|
||||
except Exception:
|
||||
return 0
|
||||
|
||||
|
||||
def _scroll_all(client, collection: str) -> List[Any]:
|
||||
pts_all = []
|
||||
offset = None
|
||||
while True:
|
||||
pts, offset = client.scroll(
|
||||
collection_name=collection,
|
||||
with_payload=True,
|
||||
with_vectors=False,
|
||||
limit=2048,
|
||||
offset=offset,
|
||||
)
|
||||
pts_all.extend(pts or [])
|
||||
if offset is None:
|
||||
break
|
||||
return pts_all
|
||||
|
||||
|
||||
def _rule_group(rule_id: str) -> str:
|
||||
if not rule_id:
|
||||
return "unknown"
|
||||
if rule_id == "callout:edge":
|
||||
return "callout"
|
||||
if rule_id.startswith("inline:"): # <—— wichtig für "inline:rel"
|
||||
return "inline"
|
||||
if rule_id.startswith("edge_defaults:"):
|
||||
return "defaults"
|
||||
if rule_id.startswith("explicit:"):
|
||||
return "explicit"
|
||||
if rule_id in ("structure:belongs_to", "structure:order"):
|
||||
return "structure"
|
||||
return "other"
|
||||
|
||||
|
||||
def main() -> None:
|
||||
cfg = QdrantConfig.from_env()
|
||||
client = get_client(cfg)
|
||||
|
||||
col_notes = f"{cfg.prefix}_notes"
|
||||
col_chunks = f"{cfg.prefix}_chunks"
|
||||
col_edges = f"{cfg.prefix}_edges"
|
||||
|
||||
# High-level counts
|
||||
notes_n = _count_collection_points(client, col_notes)
|
||||
chunks_n = _count_collection_points(client, col_chunks)
|
||||
edges_pts = _scroll_all(client, col_edges)
|
||||
edges_n = len(edges_pts)
|
||||
|
||||
# By kind / by rule group
|
||||
by_kind = Counter()
|
||||
group_counts = Counter()
|
||||
callout_buckets: Dict[Tuple[str, str], int] = defaultdict(int) # (chunk_id, kind) -> n targets
|
||||
per_note = defaultdict(lambda: {"chunks": 0, "belongs_to": 0, "next": 0, "prev": 0})
|
||||
|
||||
# Für per_note checks: chunks pro note_id aus mindnet_chunks laden
|
||||
chunks_pts = _scroll_all(client, col_chunks)
|
||||
chunks_by_note = Counter([p.payload.get("note_id") for p in chunks_pts if p.payload])
|
||||
|
||||
for p in edges_pts:
|
||||
pl = p.payload or {}
|
||||
kind = str(pl.get("kind") or pl.get("relation") or "edge")
|
||||
rule_id = str(pl.get("rule_id") or "")
|
||||
note_id = str(pl.get("note_id") or "")
|
||||
chunk_id = str(pl.get("chunk_id") or "")
|
||||
by_kind[kind] += 1
|
||||
|
||||
group = _rule_group(rule_id)
|
||||
group_counts[group] += 1
|
||||
|
||||
# Multi-Callout-Erkennung: mehrere callout-Edges gleicher Relation aus demselben Chunk
|
||||
if group == "callout" and chunk_id and kind:
|
||||
callout_buckets[(chunk_id, kind)] += 1
|
||||
|
||||
# Per-note Strukturchecks
|
||||
if note_id:
|
||||
if kind == "belongs_to":
|
||||
per_note[note_id]["belongs_to"] += 1
|
||||
elif kind == "next":
|
||||
per_note[note_id]["next"] += 1
|
||||
elif kind == "prev":
|
||||
per_note[note_id]["prev"] += 1
|
||||
|
||||
# set chunks count for per_note
|
||||
for n_id, c in chunks_by_note.items():
|
||||
per_note[n_id]["chunks"] = c
|
||||
|
||||
# final checks per note
|
||||
per_note_checks = {}
|
||||
for n_id, stats in per_note.items():
|
||||
c = stats.get("chunks", 0)
|
||||
bt = stats.get("belongs_to", 0)
|
||||
nx = stats.get("next", 0)
|
||||
pv = stats.get("prev", 0)
|
||||
per_note_checks[n_id] = {
|
||||
"chunks": c,
|
||||
"belongs_to": bt,
|
||||
"next": nx,
|
||||
"prev": pv,
|
||||
"checks": {
|
||||
"belongs_to_equals_chunks": (bt == c),
|
||||
"next_prev_match": (nx == pv == max(c - 1, 0)),
|
||||
},
|
||||
}
|
||||
|
||||
multi_callout_detected = any(v > 1 for v in callout_buckets.values())
|
||||
|
||||
out = {
|
||||
"prefix": cfg.prefix,
|
||||
"counts": {
|
||||
"notes": notes_n,
|
||||
"chunks": chunks_n,
|
||||
"edges": edges_n,
|
||||
"edges_by_kind": dict(by_kind),
|
||||
"explicit_total": group_counts.get("explicit", 0),
|
||||
"defaults_total": group_counts.get("defaults", 0),
|
||||
"callout_total": group_counts.get("callout", 0),
|
||||
"inline_total": group_counts.get("inline", 0),
|
||||
"structure_total": group_counts.get("structure", 0),
|
||||
},
|
||||
"per_note_checks": per_note_checks,
|
||||
"multi_callout_detected": bool(multi_callout_detected),
|
||||
"has_duplicates": False, # dedupe passiert beim Upsert
|
||||
}
|
||||
print(json.dumps(out, ensure_ascii=False, indent=2))
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -0,0 +1,340 @@
|
||||
#!/usr/bin/env python3
|
||||
"""
|
||||
scripts/health_check_mindnet.py
|
||||
|
||||
Health-Check für den mindnet-Retriever-/Query-Endpoint.
|
||||
|
||||
Funktion:
|
||||
- Führt POST-Requests auf /query in verschiedenen Modi aus (standard: semantic + hybrid).
|
||||
- Prüft Status-Code, JSON-Struktur und Anzahl der Treffer.
|
||||
- Kennzeichnet Probleme als:
|
||||
- status="ok"
|
||||
- status="warning" (z.B. Timeout)
|
||||
- status="error" (harte Fehler wie HTTP-Fehler, JSON-Fehler etc.)
|
||||
|
||||
Exit-Code:
|
||||
- Default (tolerant):
|
||||
- overall_status = "ok" (inkl. warnings) → Exit-Code 0
|
||||
- overall_status = "error" → Exit-Code 1
|
||||
- Mit --strict:
|
||||
- warnings werden wie errors behandelt → Exit-Code 1
|
||||
|
||||
Beispiele:
|
||||
|
||||
python scripts/health_check_mindnet.py \
|
||||
--url http://127.0.0.1:8001/query \
|
||||
--query "embeddings" \
|
||||
--top-k 3
|
||||
|
||||
python scripts/health_check_mindnet.py \
|
||||
--url http://127.0.0.1:8001/query \
|
||||
--query "embeddings" \
|
||||
--top-k 3 \
|
||||
--timeout 15 \
|
||||
--modes hybrid
|
||||
|
||||
# Strenger Modus (warnings → Exit-Code 1)
|
||||
python scripts/health_check_mindnet.py --strict
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import argparse
|
||||
import json
|
||||
import socket
|
||||
import sys
|
||||
import time
|
||||
import urllib.error
|
||||
import urllib.request
|
||||
from typing import Any, Dict, List
|
||||
|
||||
|
||||
def _post_query(
|
||||
url: str,
|
||||
mode: str,
|
||||
query: str,
|
||||
top_k: int,
|
||||
timeout: float,
|
||||
) -> Dict[str, Any]:
|
||||
"""
|
||||
Führt einen POST auf den /query-Endpoint aus und gibt das dekodierte JSON zurück.
|
||||
|
||||
Raises:
|
||||
urllib.error.URLError, urllib.error.HTTPError, RuntimeError
|
||||
"""
|
||||
payload = {
|
||||
"mode": mode,
|
||||
"query": query,
|
||||
"top_k": top_k,
|
||||
"expand": {
|
||||
"depth": 1 if mode == "hybrid" else 0,
|
||||
"edge_types": ["references", "belongs_to", "prev", "next"],
|
||||
},
|
||||
"ret": {
|
||||
"with_paths": True,
|
||||
"with_notes": True,
|
||||
"with_chunks": True,
|
||||
},
|
||||
}
|
||||
|
||||
data = json.dumps(payload).encode("utf-8")
|
||||
req = urllib.request.Request(
|
||||
url=url,
|
||||
data=data,
|
||||
headers={"Content-Type": "application/json"},
|
||||
method="POST",
|
||||
)
|
||||
|
||||
start = time.monotonic()
|
||||
with urllib.request.urlopen(req, timeout=timeout) as resp:
|
||||
status_code = resp.getcode()
|
||||
body = resp.read().decode("utf-8")
|
||||
end = time.monotonic()
|
||||
|
||||
latency_ms = int((end - start) * 1000)
|
||||
|
||||
if status_code != 200:
|
||||
raise RuntimeError(f"HTTP {status_code}: {body}")
|
||||
|
||||
try:
|
||||
doc = json.loads(body)
|
||||
except json.JSONDecodeError as exc:
|
||||
raise RuntimeError(f"Antwort ist kein gültiges JSON: {exc}") from exc
|
||||
|
||||
if not isinstance(doc, dict):
|
||||
raise RuntimeError("Antwort ist kein JSON-Objekt")
|
||||
|
||||
doc["_latency_ms"] = latency_ms
|
||||
return doc
|
||||
|
||||
|
||||
def _validate_response(
|
||||
mode: str,
|
||||
doc: Dict[str, Any],
|
||||
min_results: int,
|
||||
) -> Dict[str, Any]:
|
||||
"""
|
||||
Prüft die wichtigsten Invarianten des Query-Responses.
|
||||
|
||||
Rückgabe-Format:
|
||||
|
||||
{
|
||||
"status": "ok" | "warning" | "error",
|
||||
"message": str,
|
||||
"latency_ms": int,
|
||||
"used_mode": str,
|
||||
"result_count": int
|
||||
}
|
||||
"""
|
||||
latency_ms = int(doc.get("_latency_ms", -1))
|
||||
used_mode = str(doc.get("used_mode", "unknown"))
|
||||
results = doc.get("results", [])
|
||||
|
||||
if not isinstance(results, list):
|
||||
return {
|
||||
"status": "error",
|
||||
"message": "results ist kein Array",
|
||||
"latency_ms": latency_ms,
|
||||
"used_mode": used_mode,
|
||||
"result_count": -1,
|
||||
}
|
||||
|
||||
result_count = len(results)
|
||||
|
||||
if result_count < min_results:
|
||||
# Zu wenige Ergebnisse sind aus Sicht eines Health-Checks eher ein Error,
|
||||
# da die Retrieval-Qualität nicht gewährleistet ist.
|
||||
return {
|
||||
"status": "error",
|
||||
"message": f"zu wenige Ergebnisse: {result_count} < {min_results}",
|
||||
"latency_ms": latency_ms,
|
||||
"used_mode": used_mode,
|
||||
"result_count": result_count,
|
||||
}
|
||||
|
||||
sample_msg = "OK"
|
||||
if result_count > 0:
|
||||
first = results[0]
|
||||
if not isinstance(first, dict):
|
||||
sample_msg = "first result ist kein Objekt"
|
||||
else:
|
||||
# einfache Plausibilitätschecks (nur Zugriff, kein strikter Typcheck)
|
||||
_ = first.get("note_id")
|
||||
_ = first.get("chunk_id")
|
||||
_ = first.get("total_score")
|
||||
|
||||
status = "ok" if sample_msg == "OK" else "warning"
|
||||
return {
|
||||
"status": status,
|
||||
"message": sample_msg,
|
||||
"latency_ms": latency_ms,
|
||||
"used_mode": used_mode,
|
||||
"result_count": result_count,
|
||||
}
|
||||
|
||||
|
||||
def run_health_check(
|
||||
url: str,
|
||||
query: str,
|
||||
top_k: int,
|
||||
timeout: float,
|
||||
modes: List[str],
|
||||
min_results: int,
|
||||
) -> Dict[str, Any]:
|
||||
"""
|
||||
Führt den Health-Check für die angegebenen Modi aus.
|
||||
"""
|
||||
report: Dict[str, Any] = {
|
||||
"url": url,
|
||||
"query": query,
|
||||
"top_k": top_k,
|
||||
"timeout_s": timeout,
|
||||
"checks": [],
|
||||
"overall_status": "ok", # wird unten ggf. auf warning oder error gesetzt
|
||||
}
|
||||
|
||||
overall_status = "ok"
|
||||
|
||||
for mode in modes:
|
||||
entry: Dict[str, Any] = {
|
||||
"mode": mode,
|
||||
"status": "error",
|
||||
"message": "",
|
||||
}
|
||||
try:
|
||||
doc = _post_query(url=url, mode=mode, query=query, top_k=top_k, timeout=timeout)
|
||||
validation = _validate_response(mode=mode, doc=doc, min_results=min_results)
|
||||
entry.update(validation)
|
||||
|
||||
except urllib.error.HTTPError as exc:
|
||||
entry["status"] = "error"
|
||||
entry["message"] = f"HTTP-Fehler: {exc}"
|
||||
entry["latency_ms"] = -1
|
||||
entry["used_mode"] = mode
|
||||
entry["result_count"] = -1
|
||||
|
||||
except urllib.error.URLError as exc:
|
||||
# URLError kann u.a. socket.timeout enthalten.
|
||||
lat = -1
|
||||
entry["latency_ms"] = lat
|
||||
entry["used_mode"] = mode
|
||||
entry["result_count"] = -1
|
||||
|
||||
if isinstance(exc.reason, socket.timeout):
|
||||
entry["status"] = "warning"
|
||||
entry["message"] = f"Timeout (URLError/socket.timeout): {exc}"
|
||||
else:
|
||||
entry["status"] = "error"
|
||||
entry["message"] = f"HTTP/Netzwerkfehler: {exc}"
|
||||
|
||||
except socket.timeout as exc:
|
||||
# Direkter Timeout (falls nicht in URLError verpackt)
|
||||
entry["status"] = "warning"
|
||||
entry["message"] = f"Timeout (socket.timeout): {exc}"
|
||||
entry["latency_ms"] = -1
|
||||
entry["used_mode"] = mode
|
||||
entry["result_count"] = -1
|
||||
|
||||
except Exception as exc: # noqa: BLE001
|
||||
entry["status"] = "error"
|
||||
entry["message"] = f"unerwarteter Fehler: {exc}"
|
||||
entry["latency_ms"] = -1
|
||||
entry["used_mode"] = mode
|
||||
entry["result_count"] = -1
|
||||
|
||||
# overall_status bestimmen:
|
||||
# - error > warning > ok
|
||||
if entry["status"] == "error":
|
||||
overall_status = "error"
|
||||
elif entry["status"] == "warning" and overall_status == "ok":
|
||||
overall_status = "warning"
|
||||
|
||||
report["checks"].append(entry)
|
||||
|
||||
report["overall_status"] = overall_status
|
||||
return report
|
||||
|
||||
|
||||
def parse_args(argv: List[str]) -> argparse.Namespace:
|
||||
parser = argparse.ArgumentParser(
|
||||
description="Health-Check für mindnet /query-Retriever",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--url",
|
||||
default="http://127.0.0.1:8001/query",
|
||||
help="Vollständige URL des /query-Endpunkts (Default: %(default)s)",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--query",
|
||||
default="embeddings",
|
||||
help="Test-Query für den Health-Check (Default: %(default)s)",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--top-k",
|
||||
type=int,
|
||||
default=3,
|
||||
help="Top-K für die Testabfrage (Default: %(default)s)",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--timeout",
|
||||
type=float,
|
||||
default=5.0,
|
||||
help="Timeout in Sekunden pro Request (Default: %(default)s)",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--modes",
|
||||
nargs="+",
|
||||
default=["semantic", "hybrid"],
|
||||
help="Zu prüfende Modi (Default: semantic hybrid)",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--min-results",
|
||||
type=int,
|
||||
default=1,
|
||||
help="Minimale Anzahl erwarteter Ergebnisse (Default: %(default)s)",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--strict",
|
||||
action="store_true",
|
||||
help="Warnings als Fehler behandeln (overall_status=warning führt zu Exit-Code 1)",
|
||||
)
|
||||
return parser.parse_args(argv)
|
||||
|
||||
|
||||
def main(argv: List[str]) -> int:
|
||||
args = parse_args(argv)
|
||||
|
||||
report = run_health_check(
|
||||
url=args.url,
|
||||
query=args.query,
|
||||
top_k=args.top_k,
|
||||
timeout=args.timeout,
|
||||
modes=args.modes,
|
||||
min_results=args.min_results,
|
||||
)
|
||||
|
||||
overall = report.get("overall_status", "error")
|
||||
|
||||
# Menschlich lesbare Zusammenfassung:
|
||||
print(f"mindnet health: {overall}")
|
||||
for c in report["checks"]:
|
||||
mode = c.get("mode")
|
||||
status = c.get("status")
|
||||
latency = c.get("latency_ms")
|
||||
result_count = c.get("result_count")
|
||||
msg = c.get("message", "")
|
||||
print(f" - {mode}: {status} (latency={latency} ms, results={result_count}) {msg}")
|
||||
|
||||
# JSON-Ausgabe (z.B. für n8n):
|
||||
print(json.dumps(report, ensure_ascii=False))
|
||||
|
||||
if args.strict:
|
||||
# strict: warning wird wie error behandelt
|
||||
return 0 if overall == "ok" else 1
|
||||
else:
|
||||
# tolerant: nur echte errors führen zu Exit-Code 1
|
||||
return 0 if overall in ("ok", "warning") else 1
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
raise SystemExit(main(sys.argv[1:]))
|
||||
@@ -124,7 +124,7 @@ def load_type_registry() -> dict:
|
||||
p = _env("MINDNET_TYPES_FILE", None)
|
||||
if p and os.path.exists(p):
|
||||
return _load_json_or_yaml(p)
|
||||
fallback = os.path.abspath("./types.yaml")
|
||||
fallback = os.path.abspath("./config/types.yaml") if os.path.exists("./config/types.yaml") else os.path.abspath("./types.yaml")
|
||||
return _load_json_or_yaml(fallback)
|
||||
|
||||
def get_type_config(note_type: Optional[str], reg: dict) -> dict:
|
||||
@@ -144,19 +144,43 @@ def resolve_note_type(requested: Optional[str], reg: dict) -> str:
|
||||
return "concept" if "concept" in types else (requested or "concept")
|
||||
|
||||
def effective_chunk_profile(note_type: str, reg: dict) -> Optional[str]:
|
||||
"""Resolve chunk_profile for type or from defaults/global.
|
||||
Accepts symbolic profiles: short|medium|long|default.
|
||||
"""
|
||||
cfg = get_type_config(note_type, reg)
|
||||
prof = cfg.get("chunk_profile")
|
||||
if isinstance(prof, str) and prof in {"short", "medium", "long"}:
|
||||
prof = (cfg.get("chunk_profile") if isinstance(cfg, dict) else None)
|
||||
if isinstance(prof, str) and prof:
|
||||
return prof
|
||||
return None
|
||||
# defaults fallbacks
|
||||
for key in ("defaults", "default", "global"):
|
||||
dcfg = reg.get(key) if isinstance(reg, dict) else None
|
||||
if isinstance(dcfg, dict):
|
||||
dprof = dcfg.get("chunk_profile")
|
||||
if isinstance(dprof, str) and dprof:
|
||||
return dprof
|
||||
return "default"
|
||||
|
||||
def effective_retriever_weight(note_type: str, reg: dict) -> Optional[float]:
|
||||
"""Resolve retriever_weight for type or defaults; returns float.
|
||||
"""
|
||||
cfg = get_type_config(note_type, reg)
|
||||
w = cfg.get("retriever_weight")
|
||||
w = (cfg.get("retriever_weight") if isinstance(cfg, dict) else None)
|
||||
try:
|
||||
return float(w) if w is not None else None
|
||||
if w is not None:
|
||||
return float(w)
|
||||
except Exception:
|
||||
return None
|
||||
pass
|
||||
# defaults fallbacks
|
||||
for key in ("defaults", "default", "global"):
|
||||
dcfg = reg.get(key) if isinstance(reg, dict) else None
|
||||
if isinstance(dcfg, dict):
|
||||
dw = dcfg.get("retriever_weight")
|
||||
try:
|
||||
if dw is not None:
|
||||
return float(dw)
|
||||
except Exception:
|
||||
pass
|
||||
return 1.0
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------
|
||||
|
||||
@@ -0,0 +1,115 @@
|
||||
#!/usr/bin/env python3
|
||||
# -*- coding: utf-8 -*-
|
||||
"""
|
||||
scripts/payload_dryrun.py (zeigt, was VOR dem Upsert tatsächlich in den Payloads steht)
|
||||
- KEIN Überschreiben der Note-Payload mehr
|
||||
- types.yaml ist maßgeblich (gemäß app/core/note_payload.py & chunk_payload.py)
|
||||
"""
|
||||
from __future__ import annotations
|
||||
import argparse, os, json
|
||||
from typing import Any, Dict, List, Optional
|
||||
|
||||
from app.core.parser import read_markdown, normalize_frontmatter, validate_required_frontmatter
|
||||
from app.core.note_payload import make_note_payload
|
||||
from app.core.chunker import assemble_chunks
|
||||
from app.core.chunk_payload import make_chunk_payloads
|
||||
try:
|
||||
from app.core.derive_edges import build_edges_for_note
|
||||
except Exception:
|
||||
from app.core.edges import build_edges_for_note # type: ignore
|
||||
|
||||
def main():
|
||||
ap = argparse.ArgumentParser()
|
||||
ap.add_argument("--vault", required=True)
|
||||
ap.add_argument("--note-id")
|
||||
ap.add_argument("--with-edges", action="store_true")
|
||||
args = ap.parse_args()
|
||||
|
||||
root = os.path.abspath(args.vault)
|
||||
|
||||
files: List[str] = []
|
||||
for dp, _, fns in os.walk(root):
|
||||
for fn in fns:
|
||||
if fn.lower().endswith(".md"):
|
||||
files.append(os.path.join(dp, fn))
|
||||
files.sort()
|
||||
|
||||
for path in files:
|
||||
parsed = read_markdown(path)
|
||||
if not parsed:
|
||||
continue
|
||||
fm = normalize_frontmatter(parsed.frontmatter)
|
||||
try:
|
||||
validate_required_frontmatter(fm)
|
||||
except Exception as e:
|
||||
print(json.dumps({"path": path, "error": f"invalid frontmatter: {e}"}))
|
||||
continue
|
||||
|
||||
if args.note_id and fm.get("id") != args.note_id:
|
||||
continue
|
||||
|
||||
# Note-Payload exakt so, wie der Importer ihn baut (types.yaml maßgeblich)
|
||||
note_pl = make_note_payload(parsed,
|
||||
vault_root=root,
|
||||
hash_mode="body",
|
||||
hash_normalize="canonical",
|
||||
hash_source="parsed",
|
||||
file_path=path)
|
||||
|
||||
body_text = getattr(parsed, "body", "") or ""
|
||||
chunks = assemble_chunks(fm["id"], body_text, fm.get("type","concept"))
|
||||
|
||||
chunk_note = {
|
||||
"frontmatter": fm,
|
||||
"id": fm.get("id"),
|
||||
"type": fm.get("type"),
|
||||
"title": fm.get("title"),
|
||||
"path": note_pl.get("path") or path,
|
||||
"note_id": note_pl.get("note_id"),
|
||||
"tags": fm.get("tags"),
|
||||
}
|
||||
# make_chunk_payloads bestimmt Werte ebenfalls aus types.yaml (Frontmatter wird ignoriert)
|
||||
chunk_pls = make_chunk_payloads(
|
||||
chunk_note,
|
||||
note_pl["path"],
|
||||
chunks,
|
||||
note_text=body_text,
|
||||
types_cfg=None, # Loader aus Datei; kein Override von außen
|
||||
file_path=path,
|
||||
)
|
||||
|
||||
out = {
|
||||
"note_id": note_pl.get("note_id") or fm.get("id"),
|
||||
"title": fm.get("title"),
|
||||
"type": fm.get("type"),
|
||||
"note_payload": {
|
||||
"retriever_weight": note_pl.get("retriever_weight"),
|
||||
"chunk_profile": note_pl.get("chunk_profile")
|
||||
},
|
||||
"chunks_summary": {
|
||||
"count": len(chunk_pls),
|
||||
"first": [
|
||||
{k: chunk_pls[i].get(k) for k in ("chunk_id","index","ord","retriever_weight","chunk_profile","neighbors_prev","neighbors_next")}
|
||||
for i in range(min(3, len(chunk_pls)))
|
||||
]
|
||||
},
|
||||
"path": note_pl.get("path")
|
||||
}
|
||||
|
||||
if args.with_edges:
|
||||
edges = build_edges_for_note(
|
||||
note_id=note_pl.get("note_id") or fm.get("id"),
|
||||
chunk_payloads=chunk_pls,
|
||||
note_level_refs=note_pl.get("references") or [],
|
||||
include_note_scope_refs=False,
|
||||
)
|
||||
kinds = {}
|
||||
for e in edges:
|
||||
k = (e.get("relation") or e.get("kind") or "edge")
|
||||
kinds[k] = kinds.get(k, 0) + 1
|
||||
out["edges_summary"] = {"total": len(edges), "by_kind": kinds}
|
||||
|
||||
print(json.dumps(out, ensure_ascii=False))
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
+33
-10
@@ -2,28 +2,33 @@
|
||||
# -*- coding: utf-8 -*-
|
||||
"""
|
||||
Name: scripts/reset_qdrant.py
|
||||
Version: v1.1.1 (2025-09-05)
|
||||
Version: v1.2.0 (2025-11-11)
|
||||
Kurzbeschreibung:
|
||||
Sicheres Zurücksetzen der Qdrant-Collections für EIN Projektpräfix. Das Skript
|
||||
ermittelt zunächst die tatsächlich betroffenen Collections und zeigt eine
|
||||
Übersicht an. Anschließend wird der User um Bestätigung gebeten (interaktive
|
||||
Abfrage), bevor eine der beiden Aktionen ausgeführt wird:
|
||||
Abfrage), bevor eine der Aktionen ausgeführt wird:
|
||||
|
||||
- wipe: Löscht die Collections komplett und legt sie gemäß Projekt-Defaults neu an
|
||||
- truncate: Löscht nur alle Points (Inhalte), behält die Collection-Settings
|
||||
- wipe: Löscht die Collections komplett und legt sie gemäß Projekt-Defaults neu an.
|
||||
**Neu**: Danach werden *immer* die Payload-Indizes & Schema-Ergänzungen
|
||||
über ensure_payload_indexes() idempotent eingerichtet (falls nicht mit --no-indexes deaktiviert).
|
||||
- truncate: Löscht nur alle Points (Inhalte), behält die Collection-Settings.
|
||||
**Neu**: Optional (Default) werden im Anschluss die Payload-Indizes geprüft/ergänzt.
|
||||
|
||||
Aufruf (aus Projekt-Root, im venv):
|
||||
python3 -m scripts.reset_qdrant --mode wipe [--prefix PREFIX] [--yes] [--dry-run]
|
||||
python3 -m scripts.reset_qdrant --mode truncate [--prefix PREFIX] [--yes] [--dry-run]
|
||||
|
||||
Parameter:
|
||||
--mode wipe | truncate (Pflicht)
|
||||
--prefix Collection-Prefix (Default: env COLLECTION_PREFIX oder 'mindnet')
|
||||
--yes Keine Rückfrage, direkt ausführen (CI/CD geeignet)
|
||||
--dry-run Nur anzeigen, was passieren würde; nichts ändern
|
||||
--mode wipe | truncate (Pflicht)
|
||||
--prefix Collection-Prefix (Default: env COLLECTION_PREFIX oder 'mindnet')
|
||||
--yes Keine Rückfrage, direkt ausführen (CI/CD geeignet)
|
||||
--dry-run Nur anzeigen, was passieren würde; nichts ändern
|
||||
--no-indexes Überspringt den Schritt ensure_payload_indexes() (Standard: Indizes/Schema werden geprüft/ergänzt)
|
||||
|
||||
Umgebungsvariablen (optional):
|
||||
QDRANT_URL, QDRANT_API_KEY, COLLECTION_PREFIX, VECTOR_DIM (Default 384)
|
||||
(Zusätzliche ENV für Named Vectors etc. werden innerhalb ensure_collections/ensure_payload_indexes berücksichtigt.)
|
||||
|
||||
Sicherheitsmerkmale:
|
||||
- Betrachtet ausschließlich Collections des angegebenen Präfixes.
|
||||
@@ -34,6 +39,8 @@ Exitcodes:
|
||||
0 = OK, 1 = abgebrochen/keine Aktion, 2 = Verbindungs-/Konfigurationsfehler
|
||||
|
||||
Changelog:
|
||||
v1.2.0: ensure_payload_indexes() nach wipe/truncate standardmäßig ausführen (idempotent); --no-indexes Flag ergänzt.
|
||||
v1.1.1: Stabilisierung & Preview (2025-09-05).
|
||||
v1.1.0: Interaktive Bestätigung, --yes/--dry-run hinzugefügt, Preview der betroffenen Collections.
|
||||
v1.0.0: Wipe/Truncate ohne Bestätigungsabfrage.
|
||||
"""
|
||||
@@ -46,7 +53,13 @@ from typing import List
|
||||
from qdrant_client import QdrantClient
|
||||
from qdrant_client.http import models as rest
|
||||
|
||||
from app.core.qdrant import QdrantConfig, get_client, ensure_collections, collection_names
|
||||
from app.core.qdrant import (
|
||||
QdrantConfig,
|
||||
get_client,
|
||||
ensure_collections,
|
||||
ensure_payload_indexes,
|
||||
collection_names,
|
||||
)
|
||||
|
||||
|
||||
def resolve_existing_collections(client: QdrantClient, prefix: str) -> List[str]:
|
||||
@@ -111,12 +124,13 @@ def wipe_collections(client: QdrantClient, all_col_names: List[str], existing: L
|
||||
|
||||
|
||||
def main():
|
||||
ap = argparse.ArgumentParser(description="Wipe oder truncate mindnet-Collections in Qdrant (mit Bestätigung).")
|
||||
ap = argparse.ArgumentParser(description="Wipe oder truncate mindnet-Collections in Qdrant (mit Bestätigung & Index-Setup).")
|
||||
ap.add_argument("--mode", choices=["wipe", "truncate"], required=True,
|
||||
help="wipe = Collections löschen & neu anlegen; truncate = nur Inhalte löschen")
|
||||
ap.add_argument("--prefix", help="Collection-Prefix (Default: env COLLECTION_PREFIX oder 'mindnet')")
|
||||
ap.add_argument("--yes", action="store_true", help="Ohne Rückfrage ausführen (CI/CD)")
|
||||
ap.add_argument("--dry-run", action="store_true", help="Nur anzeigen, was passieren würde; nichts ändern")
|
||||
ap.add_argument("--no-indexes", action="store_true", help="Überspringt ensure_payload_indexes()")
|
||||
args = ap.parse_args()
|
||||
|
||||
# Qdrant-Konfiguration
|
||||
@@ -155,13 +169,22 @@ def main():
|
||||
if args.mode == "wipe":
|
||||
wipe_collections(client, all_col_names, existing)
|
||||
ensure_collections(client, cfg.prefix, cfg.dim)
|
||||
if not args.no_indexes:
|
||||
ensure_payload_indexes(client, cfg.prefix)
|
||||
print(f"Wiped & recreated (Prefix={cfg.prefix}): {all_col_names}")
|
||||
if not args.no_indexes:
|
||||
print("Payload-Indizes & Schema-Ergänzungen geprüft/ergänzt (ensure_payload_indexes).")
|
||||
else:
|
||||
if not existing:
|
||||
print("Keine existierenden Collections zum Truncaten gefunden. Beende ohne Aktion.")
|
||||
sys.exit(0)
|
||||
delete_all_points(client, existing)
|
||||
if not args.no_indexes:
|
||||
# Auch nach truncate sicherstellen, dass neue/fehlende Indizes ergänzt werden
|
||||
ensure_payload_indexes(client, cfg.prefix)
|
||||
print(f"Truncated (deleted all points in): {existing}")
|
||||
if not args.no_indexes:
|
||||
print("Payload-Indizes & Schema-Ergänzungen geprüft/ergänzt (ensure_payload_indexes).")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
|
||||
@@ -0,0 +1,116 @@
|
||||
#!/usr/bin/env python3
|
||||
# -*- coding: utf-8 -*-
|
||||
"""
|
||||
tests/assert_payload_schema.py (compat v1.1)
|
||||
|
||||
Prüft, ob die erwarteten Payload-Indizes (payload_schema) auf den drei Collections vorhanden sind.
|
||||
Kompatibel mit älteren qdrant-client Versionen (ohne with_payload_schema im Wrapper).
|
||||
|
||||
Collections & Pflichtfelder:
|
||||
- mindnet_notes : note_id, type, title, updated, tags
|
||||
- mindnet_chunks : note_id, chunk_id, index, type, tags
|
||||
- mindnet_edges : note_id, kind, scope, source_id, target_id, chunk_id
|
||||
|
||||
Nutzung:
|
||||
python3 -m tests.assert_payload_schema
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import os
|
||||
import sys
|
||||
import urllib.error
|
||||
import urllib.request
|
||||
from typing import Any, Dict, List
|
||||
|
||||
from app.core.qdrant import QdrantConfig, get_client, collection_names
|
||||
|
||||
REQUIRED = {
|
||||
"notes": ["note_id", "type", "title", "updated", "tags"],
|
||||
"chunks": ["note_id", "chunk_id", "index", "type", "tags"],
|
||||
"edges": ["note_id", "kind", "scope", "source_id", "target_id", "chunk_id"],
|
||||
}
|
||||
|
||||
def _safe_model_dump(obj: Any) -> Dict[str, Any]:
|
||||
if hasattr(obj, "model_dump"):
|
||||
return obj.model_dump()
|
||||
if hasattr(obj, "dict"):
|
||||
return obj.dict()
|
||||
return obj if isinstance(obj, dict) else {}
|
||||
|
||||
def _cfg_base_url(cfg) -> str:
|
||||
if getattr(cfg, "url", None):
|
||||
return cfg.url.rstrip('/')
|
||||
host = getattr(cfg, "host", None) or os.getenv("QDRANT_HOST") or "127.0.0.1"
|
||||
port = getattr(cfg, "port", None) or os.getenv("QDRANT_PORT") or "6333"
|
||||
return f"http://{host}:{port}"
|
||||
|
||||
def _http_get(url: str, api_key: str | None) -> Dict[str, Any]:
|
||||
req = urllib.request.Request(url, method="GET")
|
||||
if api_key:
|
||||
req.add_header("api-key", api_key)
|
||||
req.add_header("Authorization", f"Bearer {api_key}")
|
||||
req.add_header("Accept", "application/json")
|
||||
with urllib.request.urlopen(req, timeout=30) as resp:
|
||||
data = resp.read()
|
||||
return json.loads(data.decode("utf-8"))
|
||||
|
||||
def get_collection_payload_schema(client, cfg, name: str) -> Dict[str, Any] | None:
|
||||
# 1) OpenAPI client mit Flag
|
||||
try:
|
||||
oc = getattr(client, "openapi_client", None)
|
||||
if oc is not None and hasattr(oc, "collections_api"):
|
||||
api = oc.collections_api
|
||||
info = api.get_collection(collection_name=name, with_payload_schema=True)
|
||||
d = _safe_model_dump(info)
|
||||
return (d.get("result") or {}).get("payload_schema")
|
||||
except Exception:
|
||||
pass
|
||||
# 2) Wrapper (manche Versionen liefern Schema auch ohne Flag)
|
||||
try:
|
||||
info = client.get_collection(collection_name=name)
|
||||
d = _safe_model_dump(info)
|
||||
ps = (d.get("result") or {}).get("payload_schema")
|
||||
if ps is not None:
|
||||
return ps
|
||||
except Exception:
|
||||
pass
|
||||
# 3) Direkter HTTP-Call
|
||||
try:
|
||||
base = _cfg_base_url(cfg)
|
||||
url = f"{base}/collections/{name}?with_payload_schema=true"
|
||||
raw = _http_get(url, getattr(cfg, "api_key", None))
|
||||
return (raw.get("result") or {}).get("payload_schema")
|
||||
except Exception:
|
||||
return None
|
||||
|
||||
def ensure_fields(schema: Dict[str, dict] | None, required: List[str]) -> Dict[str, bool]:
|
||||
ks = set((schema or {}).keys())
|
||||
return {k: (k in ks) for k in required}
|
||||
|
||||
def main():
|
||||
cfg = QdrantConfig.from_env()
|
||||
client = get_client(cfg)
|
||||
notes, chunks, edges = collection_names(cfg.prefix)
|
||||
names = {"notes": notes, "chunks": chunks, "edges": edges}
|
||||
|
||||
report = {}
|
||||
ok_all = True
|
||||
for kind, col in names.items():
|
||||
sch = get_collection_payload_schema(client, cfg, col)
|
||||
checks = ensure_fields(sch, REQUIRED[kind])
|
||||
ok = all(checks.values())
|
||||
ok_all = ok_all and ok
|
||||
report[kind] = {
|
||||
"collection": col,
|
||||
"ok": ok,
|
||||
"missing": [k for k,v in checks.items() if not v],
|
||||
"present": [k for k,v in checks.items() if v],
|
||||
}
|
||||
|
||||
print(json.dumps({"prefix": cfg.prefix, "ok": ok_all, "report": report}, ensure_ascii=False, indent=2))
|
||||
if not ok_all:
|
||||
sys.exit(1)
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -0,0 +1,104 @@
|
||||
#!/usr/bin/env python3
|
||||
# -*- coding: utf-8 -*-
|
||||
"""
|
||||
Kleine Diagnosehilfe für mindnet_edges.
|
||||
|
||||
Nutzung (aus Projektwurzel):
|
||||
|
||||
(.venv) python tests/diag_edges_for_note.py \
|
||||
--note-id 20251130-relshow-embeddings-101
|
||||
|
||||
Optional:
|
||||
|
||||
--prefix mindnet
|
||||
|
||||
Ziel:
|
||||
- Zeigt alle Edges aus der Collection <prefix>_edges,
|
||||
die sich auf die angegebene note_id beziehen:
|
||||
* note_id == NOTE
|
||||
* oder source_id == NOTE
|
||||
* oder target_id == NOTE
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import argparse
|
||||
import os
|
||||
import sys
|
||||
from typing import List
|
||||
|
||||
# Projekt-Root (Ordner mit "app/") zum sys.path hinzufügen,
|
||||
# damit "import app.core.qdrant" funktioniert, auch wenn das Script
|
||||
# direkt aus tests/ gestartet wird.
|
||||
THIS_DIR = os.path.dirname(os.path.abspath(__file__))
|
||||
PROJECT_ROOT = os.path.dirname(THIS_DIR)
|
||||
if PROJECT_ROOT not in sys.path:
|
||||
sys.path.insert(0, PROJECT_ROOT)
|
||||
|
||||
from qdrant_client.http import models as rest # type: ignore
|
||||
|
||||
from app.core.qdrant import QdrantConfig, get_client, collection_names
|
||||
|
||||
|
||||
def fetch_edges_for_note(note_id: str, prefix: str | None, limit: int = 512) -> List[dict]:
|
||||
cfg = QdrantConfig.from_env()
|
||||
if prefix:
|
||||
cfg.prefix = prefix
|
||||
|
||||
client = get_client(cfg)
|
||||
_, _, edges_col = collection_names(cfg.prefix)
|
||||
|
||||
conditions = []
|
||||
for field in ("note_id", "source_id", "target_id"):
|
||||
conditions.append(
|
||||
rest.FieldCondition(
|
||||
key=field,
|
||||
match=rest.MatchValue(value=str(note_id)),
|
||||
)
|
||||
)
|
||||
|
||||
flt = rest.Filter(
|
||||
should=conditions,
|
||||
)
|
||||
|
||||
points, _ = client.scroll(
|
||||
collection_name=edges_col,
|
||||
scroll_filter=flt,
|
||||
with_payload=True,
|
||||
with_vectors=False,
|
||||
limit=limit,
|
||||
)
|
||||
|
||||
out: List[dict] = []
|
||||
for p in points or []:
|
||||
pl = dict(p.payload or {})
|
||||
if pl:
|
||||
out.append(pl)
|
||||
return out
|
||||
|
||||
|
||||
def main() -> None:
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument("--note-id", required=True, help="Note-ID, z. B. 20251130-relshow-embeddings-101")
|
||||
parser.add_argument("--prefix", default=None, help="Qdrant-Prefix (Default: aus ENV/QdrantConfig)")
|
||||
args = parser.parse_args()
|
||||
|
||||
note_id = args.note_id
|
||||
prefix = args.prefix
|
||||
|
||||
edges = fetch_edges_for_note(note_id, prefix)
|
||||
print(f"Prefix: {prefix or '(aus ENV)'}")
|
||||
print(f"Note-ID: {note_id}")
|
||||
print(f"Gefundene Edges: {len(edges)}")
|
||||
print("-" * 60)
|
||||
|
||||
for i, e in enumerate(edges, start=1):
|
||||
print(f"[{i}] kind={e.get('kind')} confidence={e.get('confidence')}")
|
||||
print(f" note_id = {e.get('note_id')}")
|
||||
print(f" source_id= {e.get('source_id')}")
|
||||
print(f" target_id= {e.get('target_id')}")
|
||||
print()
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -0,0 +1,56 @@
|
||||
#!/usr/bin/env python3
|
||||
# -*- coding: utf-8 -*-
|
||||
"""
|
||||
scripts/diag_payload_indexes.py (v1.1)
|
||||
|
||||
Zeigt payload_schema (Indizes) je Collection.
|
||||
WICHTIG: Einige Qdrant-Versionen liefern payload_schema nur, wenn
|
||||
`with_payload_schema=true` gesetzt wird. Daher setzen wir das Flag explizit.
|
||||
"""
|
||||
from __future__ import annotations
|
||||
import argparse, json
|
||||
from qdrant_client.http import models as rest
|
||||
from app.core.qdrant import QdrantConfig, get_client, collection_names
|
||||
|
||||
def compact_schema(ps: dict | None) -> dict:
|
||||
if not isinstance(ps, dict):
|
||||
return {}
|
||||
out = {}
|
||||
for k, v in ps.items():
|
||||
if isinstance(v, dict) and "type" in v:
|
||||
out[k] = v["type"]
|
||||
else:
|
||||
out[k] = v
|
||||
return out
|
||||
|
||||
def main():
|
||||
ap = argparse.ArgumentParser()
|
||||
ap.add_argument("--raw", action="store_true")
|
||||
args = ap.parse_args()
|
||||
|
||||
cfg = QdrantConfig.from_env()
|
||||
client = get_client(cfg)
|
||||
notes, chunks, edges = collection_names(cfg.prefix)
|
||||
|
||||
cols = [notes, chunks, edges]
|
||||
result = []
|
||||
for name in cols:
|
||||
info = client.get_collection(collection_name=name, with_payload_schema=True)
|
||||
d = info.model_dump() if hasattr(info, "model_dump") else (info.dict() if hasattr(info, "dict") else info)
|
||||
payload_schema = (d.get("result") or {}).get("payload_schema")
|
||||
vectors = (d.get("result") or {}).get("vectors")
|
||||
if isinstance(vectors, dict) and "config" in vectors:
|
||||
vectors = vectors.get("config")
|
||||
if args.raw:
|
||||
result.append({"collection": name, "raw": d})
|
||||
else:
|
||||
result.append({
|
||||
"name": name,
|
||||
"vectors": vectors,
|
||||
"payload_schema": compact_schema(payload_schema),
|
||||
"segments_count": (d.get("result") or {}).get("segments_count"),
|
||||
})
|
||||
print(json.dumps({"prefix": cfg.prefix, "collections": result}, ensure_ascii=False, indent=2))
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -0,0 +1,120 @@
|
||||
#!/usr/bin/env python3
|
||||
# -*- coding: utf-8 -*-
|
||||
"""
|
||||
Diagnose für den Graph-Adapter (Subgraph) in mindnet.
|
||||
|
||||
Nutzung (aus Projektwurzel):
|
||||
|
||||
(.venv) python tests/diag_subgraph_for_note.py \
|
||||
--note-id 20251130-relshow-embeddings-101 \
|
||||
--depth 1
|
||||
|
||||
Optional:
|
||||
|
||||
--prefix mindnet
|
||||
--edge-type references --edge-type related_to ...
|
||||
|
||||
Ziel:
|
||||
- Ruft app.core.graph_adapter.expand(...) für eine gegebene note_id auf.
|
||||
- Gibt edge_bonus und centrality_bonus für diese note_id aus.
|
||||
- Zeigt zusätzlich einige Nachbarn und deren Boni.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import argparse
|
||||
import os
|
||||
import sys
|
||||
from typing import List, Optional, Tuple
|
||||
|
||||
# sys.path anpassen, damit "import app...." aus tests/ funktioniert
|
||||
THIS_DIR = os.path.dirname(os.path.abspath(__file__))
|
||||
PROJECT_ROOT = os.path.dirname(THIS_DIR)
|
||||
if PROJECT_ROOT not in sys.path:
|
||||
sys.path.insert(0, PROJECT_ROOT)
|
||||
|
||||
from app.core.qdrant import QdrantConfig, get_client # type: ignore
|
||||
import app.core.graph_adapter as ga # type: ignore
|
||||
|
||||
|
||||
def build_subgraph_for_note(
|
||||
note_id: str,
|
||||
prefix: Optional[str],
|
||||
depth: int,
|
||||
edge_types: Optional[List[str]] = None,
|
||||
) -> Tuple[ga.Subgraph, str]:
|
||||
cfg = QdrantConfig.from_env()
|
||||
if prefix:
|
||||
cfg.prefix = prefix
|
||||
|
||||
client = get_client(cfg)
|
||||
sg = ga.expand(client, cfg.prefix, [note_id], depth=depth, edge_types=edge_types)
|
||||
return sg, cfg.prefix
|
||||
|
||||
|
||||
def main() -> None:
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument(
|
||||
"--note-id",
|
||||
required=True,
|
||||
help="Note-ID, z. B. 20251130-relshow-embeddings-101",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--depth",
|
||||
type=int,
|
||||
default=1,
|
||||
help="Expansionstiefe (Default: 1)",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--prefix",
|
||||
default=None,
|
||||
help="Qdrant-Prefix (Default: aus ENV/QdrantConfig)",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--edge-type",
|
||||
action="append",
|
||||
default=None,
|
||||
help="Edge-Typ-Filter (kann mehrfach angegeben werden), z. B. --edge-type references",
|
||||
)
|
||||
args = parser.parse_args()
|
||||
|
||||
note_id = args.note_id
|
||||
depth = max(0, int(args.depth))
|
||||
prefix = args.prefix
|
||||
edge_types = args.edge_type
|
||||
|
||||
sg, eff_prefix = build_subgraph_for_note(note_id, prefix, depth, edge_types=edge_types)
|
||||
|
||||
print(f"Prefix: {eff_prefix}")
|
||||
print(f"Note-ID (Seed): {note_id}")
|
||||
print(f"Depth: {depth}")
|
||||
print(f"Edge-Types-Filter: {edge_types or '(keiner)'}")
|
||||
print("-" * 60)
|
||||
|
||||
# Boni für die Seed-Note
|
||||
eb = sg.edge_bonus(note_id)
|
||||
cb = sg.centrality_bonus(note_id)
|
||||
print(f"edge_bonus({note_id}) = {eb:.4f}")
|
||||
print(f"centrality_bonus({note_id}) = {cb:.4f}")
|
||||
print("-" * 60)
|
||||
|
||||
# Ein paar weitere Knoten anzeigen (falls vorhanden)
|
||||
print("Beispiele weiterer Knoten mit Boni (max. 10):")
|
||||
printed = 0
|
||||
for node, edges in sg.adj.items():
|
||||
if node == note_id:
|
||||
continue
|
||||
if printed >= 10:
|
||||
break
|
||||
eb2 = sg.edge_bonus(node)
|
||||
cb2 = sg.centrality_bonus(node)
|
||||
print(f"- node={node}")
|
||||
print(f" edge_bonus={eb2:.4f} centrality={cb2:.4f}")
|
||||
printed += 1
|
||||
|
||||
if printed == 0:
|
||||
print("(keine weiteren Knoten im Subgraph registriert)")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -0,0 +1,89 @@
|
||||
#!/usr/bin/env python3
|
||||
# -*- coding: utf-8 -*-
|
||||
"""
|
||||
tests/ensure_indexes_and_show.py (compat v1.0.1)
|
||||
|
||||
Erzwingt ensure_payload_indexes(...) und zeigt danach payload_schema je Collection.
|
||||
Kompatibel mit qdrant-client Versionen ohne 'with_payload_schema' im Wrapper.
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import os
|
||||
import urllib.request
|
||||
import urllib.error
|
||||
from typing import Any, Dict
|
||||
|
||||
from app.core.qdrant import QdrantConfig, get_client, ensure_payload_indexes, collection_names
|
||||
|
||||
|
||||
def _safe_model_dump(obj: Any) -> Dict[str, Any]:
|
||||
if hasattr(obj, "model_dump"):
|
||||
return obj.model_dump()
|
||||
if hasattr(obj, "dict"):
|
||||
return obj.dict()
|
||||
return obj if isinstance(obj, dict) else {}
|
||||
|
||||
|
||||
def _cfg_base_url(cfg) -> str:
|
||||
if getattr(cfg, "url", None):
|
||||
return cfg.url.rstrip("/")
|
||||
host = getattr(cfg, "host", None) or os.getenv("QDRANT_HOST") or "127.0.0.1"
|
||||
port = getattr(cfg, "port", None) or os.getenv("QDRANT_PORT") or "6333"
|
||||
return f"http://{host}:{port}"
|
||||
|
||||
|
||||
def _http_get(url: str, api_key: str | None) -> Dict[str, Any]:
|
||||
req = urllib.request.Request(url, method="GET")
|
||||
if api_key:
|
||||
req.add_header("api-key", api_key)
|
||||
req.add_header("Authorization", f"Bearer {api_key}")
|
||||
req.add_header("Accept", "application/json")
|
||||
with urllib.request.urlopen(req, timeout=30) as resp:
|
||||
data = resp.read()
|
||||
return json.loads(data.decode("utf-8"))
|
||||
|
||||
|
||||
def get_collection_payload_schema(client, cfg, name: str) -> Dict[str, Any] | None:
|
||||
# Strategy 1: OpenAPI client with flag
|
||||
try:
|
||||
oc = getattr(client, "openapi_client", None)
|
||||
if oc is not None and hasattr(oc, "collections_api"):
|
||||
api = oc.collections_api
|
||||
info = api.get_collection(collection_name=name, with_payload_schema=True)
|
||||
d = _safe_model_dump(info)
|
||||
return (d.get("result") or {}).get("payload_schema")
|
||||
except Exception:
|
||||
pass
|
||||
# Strategy 2: wrapper
|
||||
try:
|
||||
info = client.get_collection(collection_name=name)
|
||||
d = _safe_model_dump(info)
|
||||
ps = (d.get("result") or {}).get("payload_schema")
|
||||
if ps is not None:
|
||||
return ps
|
||||
except Exception:
|
||||
pass
|
||||
# Strategy 3: direct HTTP
|
||||
try:
|
||||
base = _cfg_base_url(cfg)
|
||||
url = f"{base}/collections/{name}?with_payload_schema=true"
|
||||
raw = _http_get(url, getattr(cfg, "api_key", None))
|
||||
return (raw.get("result") or {}).get("payload_schema")
|
||||
except Exception:
|
||||
return None
|
||||
|
||||
|
||||
def main():
|
||||
cfg = QdrantConfig.from_env()
|
||||
client = get_client(cfg)
|
||||
ensure_payload_indexes(client, cfg.prefix)
|
||||
notes, chunks, edges = collection_names(cfg.prefix)
|
||||
res = {}
|
||||
for name in (notes, chunks, edges):
|
||||
res[name] = get_collection_payload_schema(client, cfg, name)
|
||||
print(json.dumps(res, ensure_ascii=False, indent=2))
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -0,0 +1,122 @@
|
||||
#!/usr/bin/env python3
|
||||
# -*- coding: utf-8 -*-
|
||||
"""
|
||||
scripts/test_retriever_smoke.py - einfacher Smoke-Test für den mindnet-Retriever
|
||||
|
||||
Zweck:
|
||||
- Startet eine Query gegen die laufende FastAPI-Instanz
|
||||
- Prüft, ob der /query-Endpunkt antwortet und sinnvolle Treffer liefert
|
||||
- Gibt Top-K-Ergebnisse inkl. Scores auf stdout aus
|
||||
|
||||
Verwendung:
|
||||
1) Stelle sicher, dass der mindnet-Service läuft, z. B.:
|
||||
uvicorn app.main:create_app --factory --reload
|
||||
|
||||
2) Führe dieses Skript aus:
|
||||
python scripts/test_retriever_smoke.py --query "karate" --mode hybrid
|
||||
|
||||
3) Optional:
|
||||
python scripts/test_retriever_smoke.py --query "embeddings" --mode semantic --top-k 5
|
||||
|
||||
Hinweis:
|
||||
- Standard-URL ist http://127.0.0.1:8000
|
||||
- Endpoint ist /query (bestehender Router), der QueryRequest erwartet
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import argparse
|
||||
import json
|
||||
import sys
|
||||
from typing import Any, Dict
|
||||
|
||||
import requests
|
||||
|
||||
|
||||
def _build_payload(args: argparse.Namespace) -> Dict[str, Any]:
|
||||
payload: Dict[str, Any] = {
|
||||
"mode": args.mode,
|
||||
"query": args.query,
|
||||
"top_k": args.top_k,
|
||||
}
|
||||
if args.expand_depth is not None and args.expand_depth > 0:
|
||||
expand: Dict[str, Any] = {"depth": args.expand_depth}
|
||||
if args.expand_edge_types:
|
||||
expand["edge_types"] = args.expand_edge_types
|
||||
payload["expand"] = expand
|
||||
return payload
|
||||
|
||||
|
||||
def main(argv: list[str] | None = None) -> int:
|
||||
parser = argparse.ArgumentParser(description="Smoke-Test für mindnet-Retriever (/query)")
|
||||
parser.add_argument("--url", default="http://127.0.0.1:8000/query", help="Basis-URL des Query-Endpunkts")
|
||||
parser.add_argument("--query", required=True, help="Text-Query für den Retriever")
|
||||
parser.add_argument("--mode", choices=["semantic", "hybrid"], default="hybrid", help="Retrieval-Modus")
|
||||
parser.add_argument("--top-k", type=int, default=5, help="Anzahl der gewünschten Treffer")
|
||||
parser.add_argument("--expand-depth", type=int, default=0, help="Edge-Expansionstiefe (nur hybrid)")
|
||||
parser.add_argument(
|
||||
"--expand-edge-types",
|
||||
nargs="*",
|
||||
default=None,
|
||||
help="Liste von Edge-Typen für die Expansion (z. B. references depends_on)",
|
||||
)
|
||||
|
||||
args = parser.parse_args(argv)
|
||||
|
||||
payload = _build_payload(args)
|
||||
|
||||
try:
|
||||
resp = requests.post(args.url, json=payload, timeout=10)
|
||||
except Exception as exc: # pragma: no cover - Netzwerkfehler
|
||||
print(f"[ERROR] HTTP-Request fehlgeschlagen: {exc}", file=sys.stderr)
|
||||
return 1
|
||||
|
||||
if resp.status_code != 200:
|
||||
print(f"[ERROR] Unerwarteter Statuscode: {resp.status_code}", file=sys.stderr)
|
||||
print(resp.text, file=sys.stderr)
|
||||
return 1
|
||||
|
||||
try:
|
||||
data = resp.json()
|
||||
except Exception as exc: # pragma: no cover - JSON-Fehler
|
||||
print(f"[ERROR] Antwort ist kein gültiges JSON: {exc}", file=sys.stderr)
|
||||
print(resp.text, file=sys.stderr)
|
||||
return 1
|
||||
|
||||
print("=== mindnet Retriever Smoke-Test ===")
|
||||
print(f"URL: {args.url}")
|
||||
print(f"Mode: {args.mode}")
|
||||
print(f"Query: {args.query!r}")
|
||||
print(f"Top-K: {args.top_k}")
|
||||
print()
|
||||
|
||||
results = data.get("results") or []
|
||||
used_mode = data.get("used_mode")
|
||||
latency_ms = data.get("latency_ms")
|
||||
|
||||
print(f"used_mode = {used_mode}, latency = {latency_ms} ms")
|
||||
print(f"Treffer insgesamt: {len(results)}")
|
||||
print()
|
||||
|
||||
for i, hit in enumerate(results, start=1):
|
||||
note_id = hit.get("note_id")
|
||||
node_id = hit.get("node_id")
|
||||
total = hit.get("total_score")
|
||||
sem = hit.get("semantic_score")
|
||||
edge_bonus = hit.get("edge_bonus")
|
||||
cent_bonus = hit.get("centrality_bonus")
|
||||
src = hit.get("source") or {}
|
||||
path = src.get("path")
|
||||
section = src.get("section")
|
||||
|
||||
print(f"[{i}] note_id={note_id} node_id={node_id}")
|
||||
print(f" total={total:.4f} (semantic={sem:.4f} edge={edge_bonus:.4f} centrality={cent_bonus:.4f})")
|
||||
if path or section:
|
||||
print(f" source: {path or '-'} :: {section or '-'}")
|
||||
print()
|
||||
|
||||
return 0
|
||||
|
||||
|
||||
if __name__ == "__main__": # pragma: no cover
|
||||
raise SystemExit(main())
|
||||
@@ -0,0 +1,71 @@
|
||||
#!/usr/bin/env python3
|
||||
# -*- coding: utf-8 -*-
|
||||
"""
|
||||
scripts/show_edges_for_note.py
|
||||
|
||||
Zeigt Kanten einer Note (gefiltert nach kind/scope) in einer kompakten Form.
|
||||
Aufrufbeispiele:
|
||||
python3 -m scripts.show_edges_for_note --note-id 20251110-ollama-llm-9f0a12 --kinds references,next,prev --limit 10
|
||||
python3 -m scripts.show_edges_for_note --title "Qdrant Vektordatenbank" --scope note
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
import argparse, json, os, sys
|
||||
from typing import Dict, Any, List, Tuple
|
||||
from qdrant_client.http import models as rest
|
||||
|
||||
from app.core.qdrant import QdrantConfig, get_client
|
||||
|
||||
def collections(prefix: str) -> Tuple[str, str, str]:
|
||||
return f"{prefix}_notes", f"{prefix}_chunks", f"{prefix}_edges"
|
||||
|
||||
def find_note_by_title(client, prefix: str, title: str) -> str | None:
|
||||
notes_col, _, _ = collections(prefix)
|
||||
f = rest.Filter(must=[rest.FieldCondition(key="title", match=rest.MatchText(text=title))])
|
||||
pts, _ = client.scroll(collection_name=notes_col, scroll_filter=f, with_payload=True, with_vectors=False, limit=1)
|
||||
if not pts:
|
||||
return None
|
||||
return pts[0].payload.get("note_id")
|
||||
|
||||
def fetch_edges_for_note(client, prefix: str, note_id: str, kinds: List[str] | None, scope: str | None, limit: int) -> List[Dict[str, Any]]:
|
||||
_, _, edges_col = collections(prefix)
|
||||
must = [rest.FieldCondition(key="note_id", match=rest.MatchValue(value=note_id))]
|
||||
if scope:
|
||||
must.append(rest.FieldCondition(key="scope", match=rest.MatchValue(value=scope)))
|
||||
if kinds:
|
||||
must.append(rest.FieldCondition(key="kind", match=rest.MatchAny(any=kinds)))
|
||||
f = rest.Filter(must=must)
|
||||
pts, _ = client.scroll(collection_name=edges_col, scroll_filter=f, with_payload=True, with_vectors=False, limit=limit)
|
||||
return [p.payload for p in pts]
|
||||
|
||||
def main():
|
||||
ap = argparse.ArgumentParser()
|
||||
ap.add_argument("--note-id")
|
||||
ap.add_argument("--title")
|
||||
ap.add_argument("--kinds", help="CSV: references,next,prev,belongs_to,backlink")
|
||||
ap.add_argument("--scope", choices=["note","chunk"])
|
||||
ap.add_argument("--limit", type=int, default=25)
|
||||
args = ap.parse_args()
|
||||
|
||||
cfg = QdrantConfig.from_env()
|
||||
client = get_client(cfg)
|
||||
|
||||
nid = args.note_id
|
||||
if not nid and args.title:
|
||||
nid = find_note_by_title(client, cfg.prefix, args.title)
|
||||
if not nid:
|
||||
print(json.dumps({"error": f"note with title '{args.title}' not found"}))
|
||||
sys.exit(2)
|
||||
if not nid:
|
||||
print(json.dumps({"error": "please provide --note-id or --title"}))
|
||||
sys.exit(2)
|
||||
|
||||
kinds = None
|
||||
if args.kinds:
|
||||
kinds = [s.strip() for s in args.kinds.split(",") if s.strip()]
|
||||
|
||||
edges = fetch_edges_for_note(client, cfg.prefix, nid, kinds, args.scope, args.limit)
|
||||
print(json.dumps({"note_id": nid, "count": len(edges), "edges": edges}, ensure_ascii=False, indent=2))
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -0,0 +1,116 @@
|
||||
#!/usr/bin/env python3
|
||||
# -*- coding: utf-8 -*-
|
||||
"""
|
||||
tests/test_edges_all.py
|
||||
Ein knapper Integrationscheck:
|
||||
- Es existieren Notes/Chunks/Edges
|
||||
- Inline-Edges (rule_id startswith "inline:") werden erkannt
|
||||
- Callout-Edges (rule_id == "callout:edge") werden erkannt
|
||||
- Defaults (rule_id startswith "edge_defaults:") werden erkannt
|
||||
- Strukturkanten stimmen (belongs_to == chunks; next == prev == chunks-1)
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
from collections import Counter, defaultdict
|
||||
from typing import Dict, Any, List, Tuple
|
||||
|
||||
from app.core.qdrant import QdrantConfig, get_client
|
||||
|
||||
|
||||
def _scroll_all(client, collection: str):
|
||||
pts_all = []
|
||||
offset = None
|
||||
while True:
|
||||
pts, offset = client.scroll(
|
||||
collection_name=collection,
|
||||
with_payload=True,
|
||||
with_vectors=False,
|
||||
limit=2048,
|
||||
offset=offset,
|
||||
)
|
||||
pts_all.extend(pts or [])
|
||||
if offset is None:
|
||||
break
|
||||
return pts_all
|
||||
|
||||
|
||||
def _rule_group(rule_id: str) -> str:
|
||||
if not rule_id:
|
||||
return "unknown"
|
||||
if rule_id == "callout:edge":
|
||||
return "callout"
|
||||
if rule_id.startswith("inline:"): # <—— wichtig für inline:rel
|
||||
return "inline"
|
||||
if rule_id.startswith("edge_defaults:"):
|
||||
return "defaults"
|
||||
if rule_id.startswith("explicit:"):
|
||||
return "explicit"
|
||||
if rule_id in ("structure:belongs_to", "structure:order"):
|
||||
return "structure"
|
||||
return "other"
|
||||
|
||||
|
||||
def main() -> None:
|
||||
cfg = QdrantConfig.from_env()
|
||||
client = get_client(cfg)
|
||||
|
||||
col_notes = f"{cfg.prefix}_notes"
|
||||
col_chunks = f"{cfg.prefix}_chunks"
|
||||
col_edges = f"{cfg.prefix}_edges"
|
||||
|
||||
notes_n = client.count(collection_name=col_notes, exact=True).count
|
||||
chunks_pts = _scroll_all(client, col_chunks)
|
||||
edges_pts = _scroll_all(client, col_edges)
|
||||
|
||||
ok = True
|
||||
|
||||
# Basisbedingungen
|
||||
if notes_n == 0 or len(chunks_pts) == 0 or len(edges_pts) == 0:
|
||||
ok = False
|
||||
|
||||
# Gruppen zählen
|
||||
g = Counter(_rule_group((p.payload or {}).get("rule_id", "")) for p in edges_pts)
|
||||
structure = g.get("structure", 0)
|
||||
explicit = g.get("explicit", 0)
|
||||
inline = g.get("inline", 0)
|
||||
callout = g.get("callout", 0)
|
||||
defaults = g.get("defaults", 0)
|
||||
|
||||
if structure == 0:
|
||||
ok = False
|
||||
# mindestens eine der expliziten Varianten vorhanden
|
||||
if (explicit + inline + callout) == 0:
|
||||
ok = False
|
||||
# defaults dürfen 0 sein, wenn types.yaml keine edge_defaults liefert – daher nur Info
|
||||
|
||||
# per-note checks
|
||||
chunks_by_note = Counter([p.payload.get("note_id") for p in chunks_pts if p.payload])
|
||||
belongs = Counter(
|
||||
(p.payload or {}).get("note_id")
|
||||
for p in edges_pts
|
||||
if (p.payload or {}).get("kind") == "belongs_to"
|
||||
)
|
||||
nxt = Counter(
|
||||
(p.payload or {}).get("note_id")
|
||||
for p in edges_pts
|
||||
if (p.payload or {}).get("kind") == "next"
|
||||
)
|
||||
prv = Counter(
|
||||
(p.payload or {}).get("note_id")
|
||||
for p in edges_pts
|
||||
if (p.payload or {}).get("kind") == "prev"
|
||||
)
|
||||
|
||||
for n_id, c in chunks_by_note.items():
|
||||
if belongs.get(n_id, 0) != c:
|
||||
ok = False
|
||||
if (nxt.get(n_id, 0) != max(c - 1, 0)) or (prv.get(n_id, 0) != max(c - 1, 0)):
|
||||
ok = False
|
||||
|
||||
print(json.dumps({"ok": ok, "notes_checked": len(chunks_by_note)}, ensure_ascii=False))
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -0,0 +1,25 @@
|
||||
#!/usr/bin/env python3
|
||||
# -*- coding: utf-8 -*-
|
||||
"""
|
||||
tests/test_edges_defaults_smoke.py
|
||||
Aggregiert Edges je Relation (kind) in Qdrant zur schnellen Sichtprüfung.
|
||||
Zeigt, ob typenbasierte edge_defaults-Relationen (z. B. depends_on/related_to) auftauchen.
|
||||
"""
|
||||
from __future__ import annotations
|
||||
from collections import Counter
|
||||
import json
|
||||
from app.core.qdrant import QdrantConfig, get_client
|
||||
|
||||
def main():
|
||||
cfg = QdrantConfig.from_env()
|
||||
cl = get_client(cfg)
|
||||
col = f"{cfg.prefix}_edges"
|
||||
pts, _ = cl.scroll(collection_name=col, with_payload=True, with_vectors=False, limit=10000)
|
||||
by_rel = Counter()
|
||||
for p in pts:
|
||||
rel = p.payload.get("kind") or p.payload.get("relation") or "edge"
|
||||
by_rel[rel] += 1
|
||||
print(json.dumps({"prefix": cfg.prefix, "relations": dict(by_rel)}, ensure_ascii=False))
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -0,0 +1,133 @@
|
||||
#!/usr/bin/env python3
|
||||
# -*- coding: utf-8 -*-
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import os
|
||||
from collections import Counter, defaultdict
|
||||
|
||||
from qdrant_client.http import models as rest
|
||||
from app.core.qdrant import QdrantConfig, get_client
|
||||
|
||||
|
||||
def _rel(pl: dict) -> str:
|
||||
return pl.get("relation") or pl.get("kind") or "edge"
|
||||
|
||||
|
||||
def _scroll(client, col):
|
||||
pts = []
|
||||
next_page = None
|
||||
while True:
|
||||
res, next_page = client.scroll(
|
||||
collection_name=col,
|
||||
with_payload=True,
|
||||
with_vectors=False,
|
||||
limit=1024,
|
||||
offset=next_page,
|
||||
)
|
||||
pts.extend(res)
|
||||
if next_page is None:
|
||||
break
|
||||
return pts
|
||||
|
||||
|
||||
def main():
|
||||
cfg = QdrantConfig.from_env()
|
||||
client = get_client(cfg)
|
||||
prefix = os.environ.get("COLLECTION_PREFIX", cfg.prefix)
|
||||
|
||||
cols = {
|
||||
"notes": f"{prefix}_notes",
|
||||
"chunks": f"{prefix}_chunks",
|
||||
"edges": f"{prefix}_edges",
|
||||
}
|
||||
|
||||
# Index: notes -> title/type
|
||||
notes_meta = {}
|
||||
for p in _scroll(client, cols["notes"]):
|
||||
pl = p.payload or {}
|
||||
nid = pl.get("note_id")
|
||||
if nid:
|
||||
notes_meta[nid] = {
|
||||
"title": pl.get("title", ""),
|
||||
"type": pl.get("type", ""),
|
||||
}
|
||||
|
||||
# chunks je note
|
||||
chunks_by_note = defaultdict(int)
|
||||
for p in _scroll(client, cols["chunks"]):
|
||||
pl = p.payload or {}
|
||||
nid = pl.get("note_id")
|
||||
if nid:
|
||||
chunks_by_note[nid] += 1
|
||||
|
||||
# edges je note
|
||||
edges_by_note = defaultdict(list)
|
||||
edges_all = _scroll(client, cols["edges"])
|
||||
for p in edges_all:
|
||||
pl = p.payload or {}
|
||||
nid = pl.get("note_id")
|
||||
if nid:
|
||||
edges_by_note[nid].append(pl)
|
||||
|
||||
# pro note ausgeben
|
||||
summary_edges = Counter()
|
||||
total_chunks = 0
|
||||
for nid in sorted(notes_meta.keys()):
|
||||
meta = notes_meta[nid]
|
||||
chunks = chunks_by_note.get(nid, 0)
|
||||
total_chunks += chunks
|
||||
|
||||
kinds = Counter(_rel(pl) for pl in edges_by_note[nid])
|
||||
summary_edges.update(kinds)
|
||||
|
||||
row = {
|
||||
"note_id": nid,
|
||||
"title": meta["title"],
|
||||
"type": meta["type"],
|
||||
"chunks": chunks,
|
||||
"edges_by_kind": dict(kinds),
|
||||
"checks": {
|
||||
"belongs_to_equals_chunks": (kinds.get("belongs_to", 0) == chunks),
|
||||
"next_prev_match": (kinds.get("next", 0) == kinds.get("prev", 0) == max(0, chunks - 1)),
|
||||
"no_duplicate_edges": _no_dupes(edges_by_note[nid]),
|
||||
},
|
||||
}
|
||||
print(json.dumps(row, ensure_ascii=False))
|
||||
|
||||
# Gesamtsummary
|
||||
total_notes = len(notes_meta)
|
||||
out = {
|
||||
"prefix": prefix,
|
||||
"summary": {
|
||||
"notes": total_notes,
|
||||
"chunks": total_chunks,
|
||||
"belongs_to": summary_edges.get("belongs_to", 0),
|
||||
"next": summary_edges.get("next", 0),
|
||||
"prev": summary_edges.get("prev", 0),
|
||||
"refs_chunk": summary_edges.get("references", 0),
|
||||
"refs_note": summary_edges.get("references_note", 0),
|
||||
"backlink": summary_edges.get("backlink", 0),
|
||||
"dup_edges": 0, # per-Note geprüft
|
||||
},
|
||||
}
|
||||
print(json.dumps(out, ensure_ascii=False))
|
||||
|
||||
|
||||
def _no_dupes(pls):
|
||||
seen = set()
|
||||
for pl in pls:
|
||||
key = (
|
||||
str(pl.get("source_id") or ""),
|
||||
str(pl.get("target_id") or ""),
|
||||
str(pl.get("relation") or pl.get("kind") or ""),
|
||||
str(pl.get("rule_id") or ""),
|
||||
)
|
||||
if key in seen:
|
||||
return False
|
||||
seen.add(key)
|
||||
return True
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -0,0 +1,74 @@
|
||||
#!/usr/bin/env python3
|
||||
# -*- coding: utf-8 -*-
|
||||
"""
|
||||
scripts/test_edges_smoke_fast.py — Zählung über /count (sehr schnell)
|
||||
|
||||
Verwendet Qdrant 'count' API je Note/Kind (anstatt scroll), dadurch sehr schnelle Ausführung.
|
||||
Optionen:
|
||||
--max-notes N : prüft nur die ersten N Notizen
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
import argparse, json
|
||||
from typing import Dict, Any, List, Tuple
|
||||
from qdrant_client.http import models as rest
|
||||
|
||||
from app.core.qdrant import QdrantConfig, get_client
|
||||
|
||||
KINDS = ["belongs_to", "next", "prev", "references", "backlink"]
|
||||
|
||||
def collections(prefix: str) -> Tuple[str, str, str]:
|
||||
return f"{prefix}_notes", f"{prefix}_chunks", f"{prefix}_edges"
|
||||
|
||||
def list_note_ids(client, notes_col: str, max_notes: int | None) -> List[Dict[str, Any]]:
|
||||
pts, _ = client.scroll(collection_name=notes_col, with_payload=True, with_vectors=False, limit=max_notes or 1024)
|
||||
out = []
|
||||
for p in pts or []:
|
||||
pl = p.payload or {}
|
||||
nid = pl.get("note_id") or pl.get("id")
|
||||
if nid:
|
||||
out.append({"note_id": nid, "title": pl.get("title"), "type": pl.get("type")})
|
||||
return out
|
||||
|
||||
def count_points(client, col: str, filt: rest.Filter) -> int:
|
||||
res = client.count(collection_name=col, count_filter=filt, exact=True)
|
||||
return int(getattr(res, "count", 0))
|
||||
|
||||
def main():
|
||||
ap = argparse.ArgumentParser()
|
||||
ap.add_argument("--max-notes", type=int)
|
||||
args = ap.parse_args()
|
||||
|
||||
cfg = QdrantConfig.from_env()
|
||||
client = get_client(cfg)
|
||||
notes_col, chunks_col, edges_col = collections(cfg.prefix)
|
||||
notes = list_note_ids(client, notes_col, args.max_notes)
|
||||
|
||||
summary = {"notes": 0, "chunks": 0, "edges": 0}
|
||||
for n in notes:
|
||||
nid = n["note_id"]
|
||||
summary["notes"] += 1
|
||||
|
||||
filt_note = rest.Filter(must=[rest.FieldCondition(key="note_id", match=rest.MatchValue(value=nid))])
|
||||
chunk_cnt = count_points(client, chunks_col, filt_note)
|
||||
summary["chunks"] += chunk_cnt
|
||||
|
||||
# counts per kind (edges)
|
||||
by_kind: Dict[str, int] = {}
|
||||
for k in KINDS:
|
||||
f = rest.Filter(must=[
|
||||
rest.FieldCondition(key="note_id", match=rest.MatchValue(value=nid)),
|
||||
rest.FieldCondition(key="kind", match=rest.MatchValue(value=k)),
|
||||
])
|
||||
c = count_points(client, edges_col, f)
|
||||
if c:
|
||||
by_kind[k] = c
|
||||
|
||||
summary["edges"] += sum(by_kind.values())
|
||||
line = {"note_id": nid, "title": n.get("title"), "type": n.get("type"), "chunks": chunk_cnt, "edges_by_kind": by_kind}
|
||||
print(json.dumps(line, ensure_ascii=False))
|
||||
|
||||
print(json.dumps({"prefix": cfg.prefix, "summary": summary}, ensure_ascii=False))
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -0,0 +1,67 @@
|
||||
import argparse
|
||||
import requests
|
||||
import json
|
||||
import sys
|
||||
|
||||
def main():
|
||||
parser = argparse.ArgumentParser(description="Testet die Explanation-Funktion des Retrievers.")
|
||||
parser.add_argument("--url", default="http://localhost:8002/query", help="URL zum Query Endpoint")
|
||||
parser.add_argument("--query", default="mindnet", help="Suchbegriff")
|
||||
args = parser.parse_args()
|
||||
|
||||
payload = {
|
||||
"mode": "hybrid",
|
||||
"query": args.query,
|
||||
"top_k": 1,
|
||||
"explain": True, # Das neue Flag aus WP-04b
|
||||
"expand": {
|
||||
"depth": 1,
|
||||
"edge_types": ["references", "belongs_to", "depends_on", "related_to"]
|
||||
}
|
||||
}
|
||||
|
||||
print(f"--- Sende Request an {args.url} ---")
|
||||
try:
|
||||
r = requests.post(args.url, json=payload)
|
||||
r.raise_for_status()
|
||||
data = r.json()
|
||||
except Exception as e:
|
||||
print(f"FEHLER: {e}")
|
||||
sys.exit(1)
|
||||
|
||||
hits = data.get("results", [])
|
||||
if not hits:
|
||||
print("Keine Treffer gefunden.")
|
||||
return
|
||||
|
||||
top_hit = hits[0]
|
||||
print(f"\nTop Treffer: {top_hit.get('node_id')}")
|
||||
print(f"Total Score: {top_hit.get('total_score'):.4f}")
|
||||
|
||||
# Explanation prüfen
|
||||
expl = top_hit.get("explanation")
|
||||
if not expl:
|
||||
print("\n[FAIL] Keine 'explanation' im Response erhalten!")
|
||||
sys.exit(1)
|
||||
|
||||
print("\n--- EXPLANATION LAYER ---")
|
||||
|
||||
bd = expl.get("breakdown", {})
|
||||
print(f"Breakdown:")
|
||||
print(f" Semantic Contribution: {bd.get('semantic_contribution', 0):.4f}")
|
||||
print(f" Edge Contribution: {bd.get('edge_contribution', 0):.4f}")
|
||||
print(f" Centrality Contribution: {bd.get('centrality_contribution', 0):.4f}")
|
||||
print(f" Typ-Gewicht: {bd.get('node_weight', 0):.2f}")
|
||||
|
||||
print("\nReasons:")
|
||||
for reason in expl.get("reasons", []):
|
||||
kind = reason.get("kind", "").upper()
|
||||
msg = reason.get("message", "")
|
||||
impact = reason.get("score_impact")
|
||||
impact_str = f" ({impact:+.4f})" if impact is not None else ""
|
||||
print(f" [{kind}] {msg}{impact_str}")
|
||||
|
||||
print("\n[SUCCESS] Explanation Layer antwortet korrekt.")
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -0,0 +1,64 @@
|
||||
from app.core import retriever as r
|
||||
from app.models.dto import QueryRequest
|
||||
import app.core.qdrant as qdr
|
||||
import app.core.qdrant_points as qp
|
||||
import app.services.embeddings_client as ec
|
||||
|
||||
|
||||
class _DummyClient:
|
||||
"""Minimaler Platzhalter für QdrantClient in Unit-Tests."""
|
||||
pass
|
||||
|
||||
|
||||
def _fake_get_client(cfg):
|
||||
return _DummyClient()
|
||||
|
||||
|
||||
def _fake_embed_text(text: str):
|
||||
# Liefert stabilen Vektor; Größe ist für diesen Test egal,
|
||||
# da search_chunks_by_vector ebenfalls gefaked wird.
|
||||
return [0.0] * 16
|
||||
|
||||
|
||||
def _fake_search_chunks_by_vector(client, prefix, vector, top=10, filters=None):
|
||||
# einfache Hitliste in absteigender Score-Reihenfolge
|
||||
return [
|
||||
("chunk:1", 0.9, {"note_id": "note:1", "path": "a.md", "section_title": "S1"}),
|
||||
("chunk:2", 0.7, {"note_id": "note:2", "path": "b.md", "section_title": "S2"}),
|
||||
("chunk:3", 0.4, {"note_id": "note:3", "path": "c.md", "section_title": "S3"}),
|
||||
]
|
||||
|
||||
|
||||
def test_semantic_retrieve_basic(monkeypatch):
|
||||
# Qdrant- und Embedding-Aufrufe faken
|
||||
monkeypatch.setattr(qdr, "get_client", _fake_get_client)
|
||||
monkeypatch.setattr(ec, "embed_text", _fake_embed_text)
|
||||
monkeypatch.setattr(qp, "search_chunks_by_vector", _fake_search_chunks_by_vector)
|
||||
|
||||
req = QueryRequest(mode="semantic", query="karate trainingsplan", top_k=2)
|
||||
resp = r.semantic_retrieve(req)
|
||||
|
||||
assert resp.used_mode == "semantic"
|
||||
assert len(resp.results) == 2
|
||||
|
||||
# Scores absteigend sortiert
|
||||
scores = [h.total_score for h in resp.results]
|
||||
assert scores[0] >= scores[1]
|
||||
# Korrekte Zuordnung von note_id und Quelle
|
||||
assert resp.results[0].note_id == "note:1"
|
||||
assert resp.results[0].source["path"] == "a.md"
|
||||
|
||||
|
||||
def test_hybrid_retrieve_basic(monkeypatch):
|
||||
# Qdrant- und Embedding-Aufrufe faken
|
||||
monkeypatch.setattr(qdr, "get_client", _fake_get_client)
|
||||
monkeypatch.setattr(qp, "search_chunks_by_vector", _fake_search_chunks_by_vector)
|
||||
|
||||
# Im Hybrid-Modus arbeiten wir im Step-1-Stand nur mit query_vector
|
||||
req = QueryRequest(mode="hybrid", query_vector=[0.0] * 16, top_k=2)
|
||||
resp = r.hybrid_retrieve(req)
|
||||
|
||||
assert resp.used_mode == "hybrid"
|
||||
assert len(resp.results) == 2
|
||||
scores = [h.total_score for h in resp.results]
|
||||
assert scores[0] >= scores[1]
|
||||
@@ -0,0 +1,74 @@
|
||||
from app.core import retriever as r
|
||||
from app.models.dto import QueryRequest
|
||||
import app.core.qdrant as qdr
|
||||
import app.core.qdrant_points as qp
|
||||
import app.core.graph_adapter as ga
|
||||
import app.services.embeddings_client as ec
|
||||
|
||||
|
||||
class _DummyClient:
|
||||
pass
|
||||
|
||||
|
||||
def _fake_get_client(cfg):
|
||||
return _DummyClient()
|
||||
|
||||
|
||||
def _fake_embed_text(text: str):
|
||||
# Dimension ist egal, da wir search_chunks_by_vector faken
|
||||
return [0.0] * 384
|
||||
|
||||
|
||||
def _fake_search_chunks_by_vector(client, prefix, vector, top=10, filters=None):
|
||||
# Drei Treffer mit identischem semantic_score und retriever_weight,
|
||||
# aber unterschiedlichen Graph-Boni
|
||||
return [
|
||||
("chunk:1", 0.8, {"note_id": "note:1", "chunk_id": "c1", "path": "a.md", "section_title": "S1"}),
|
||||
("chunk:2", 0.8, {"note_id": "note:2", "chunk_id": "c2", "path": "b.md", "section_title": "S2"}),
|
||||
("chunk:3", 0.8, {"note_id": "note:3", "chunk_id": "c3", "path": "c.md", "section_title": "S3"}),
|
||||
]
|
||||
|
||||
|
||||
class _DummySubgraph:
|
||||
def __init__(self, edge_scores, cent_scores):
|
||||
self._edge = edge_scores
|
||||
self._cent = cent_scores
|
||||
|
||||
def edge_bonus(self, node_id: str) -> float:
|
||||
return self._edge.get(node_id, 0.0)
|
||||
|
||||
def centrality_bonus(self, node_id: str) -> float:
|
||||
return self._cent.get(node_id, 0.0)
|
||||
|
||||
|
||||
def _fake_expand(client, prefix, seeds, depth=1, edge_types=None):
|
||||
# Wir gewichten c2 > c1 > c3
|
||||
edge_scores = {"c1": 0.1, "c2": 0.5, "c3": 0.0}
|
||||
cent_scores = {"c1": 0.05, "c2": 0.2, "c3": 0.0}
|
||||
return _DummySubgraph(edge_scores, cent_scores)
|
||||
|
||||
|
||||
def test_hybrid_retrieve_uses_edge_scores(monkeypatch):
|
||||
# Qdrant-Client, Embeddings, Chunk-Suche und Graph-Expansion faken
|
||||
monkeypatch.setattr(qdr, "get_client", _fake_get_client)
|
||||
monkeypatch.setattr(ec, "embed_text", _fake_embed_text)
|
||||
monkeypatch.setattr(qp, "search_chunks_by_vector", _fake_search_chunks_by_vector)
|
||||
monkeypatch.setattr(ga, "expand", _fake_expand)
|
||||
|
||||
# expand.depth > 0 triggern, edge_types sind hier egal
|
||||
req = QueryRequest(mode="hybrid", query="karate", top_k=3, expand={"depth": 1, "edge_types": ["references"]})
|
||||
resp = r.hybrid_retrieve(req)
|
||||
|
||||
assert len(resp.results) == 3
|
||||
|
||||
# Da alle semantic_scores und retriever_weight gleich sind,
|
||||
# bestimmt die Summe aus edge_bonus + centrality_bonus das Ranking.
|
||||
first = resp.results[0]
|
||||
second = resp.results[1]
|
||||
third = resp.results[2]
|
||||
|
||||
assert first.note_id == "note:2" # c2 hat höchste Graph-Boni
|
||||
assert second.note_id == "note:1"
|
||||
assert third.note_id == "note:3"
|
||||
|
||||
assert first.total_score > second.total_score > third.total_score
|
||||
@@ -0,0 +1,120 @@
|
||||
#!/usr/bin/env python3
|
||||
# -*- coding: utf-8 -*-
|
||||
#
|
||||
# test_retriever_smoke.py – End-to-End Smoke-Test für den mindnet Retriever
|
||||
#
|
||||
# Dieser Smoke-Test ist für manuelle End-to-End-Tests gedacht.
|
||||
# Er wird NICHT automatisch von pytest ausgeführt.
|
||||
#
|
||||
# Empfohlener Speicherort: tests/test_retriever_smoke.py
|
||||
#
|
||||
# Aufrufbeispiele:
|
||||
#
|
||||
# python tests/test_retriever_smoke.py --query "embeddings" --mode semantic
|
||||
#
|
||||
# python tests/test_retriever_smoke.py \
|
||||
# --query "embeddings" \
|
||||
# --mode hybrid \
|
||||
# --expand-depth 1 \
|
||||
# --expand-edge-types references depends_on
|
||||
#
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import argparse
|
||||
import sys
|
||||
from typing import Any, Dict
|
||||
|
||||
import requests
|
||||
|
||||
|
||||
def _build_payload(args: argparse.Namespace) -> Dict[str, Any]:
|
||||
payload: Dict[str, Any] = {
|
||||
"mode": args.mode,
|
||||
"query": args.query,
|
||||
"top_k": args.top_k,
|
||||
}
|
||||
if args.expand_depth is not None and args.expand_depth > 0:
|
||||
expand: Dict[str, Any] = {"depth": args.expand_depth}
|
||||
if args.expand_edge_types:
|
||||
expand["edge_types"] = args.expand_edge_types
|
||||
payload["expand"] = expand
|
||||
return payload
|
||||
|
||||
|
||||
def main(argv: list[str] | None = None) -> int:
|
||||
parser = argparse.ArgumentParser(description="Smoke-Test für mindnet-Retriever (/query)")
|
||||
parser.add_argument("--url", default="http://127.0.0.1:8000/query", help="URL des Query-Endpunkts")
|
||||
parser.add_argument("--query", required=True, help="Text-Query")
|
||||
parser.add_argument("--mode", choices=["semantic", "hybrid"], default="hybrid")
|
||||
parser.add_argument("--top-k", type=int, default=5)
|
||||
parser.add_argument("--expand-depth", type=int, default=0)
|
||||
parser.add_argument(
|
||||
"--expand-edge-types",
|
||||
nargs="*",
|
||||
default=None,
|
||||
help="Edge-Typen (z. B. references depends_on)",
|
||||
)
|
||||
|
||||
args = parser.parse_args(argv)
|
||||
payload = _build_payload(args)
|
||||
|
||||
try:
|
||||
resp = requests.post(args.url, json=payload, timeout=10)
|
||||
except Exception as exc:
|
||||
print(f"[ERROR] HTTP-Request fehlgeschlagen: {exc}", file=sys.stderr)
|
||||
return 1
|
||||
|
||||
if resp.status_code != 200:
|
||||
print(f"[ERROR] Statuscode: {resp.status_code}", file=sys.stderr)
|
||||
print(resp.text, file=sys.stderr)
|
||||
return 1
|
||||
|
||||
try:
|
||||
data = resp.json()
|
||||
except Exception:
|
||||
print("[ERROR] Ungültiges JSON:", resp.text, file=sys.stderr)
|
||||
return 1
|
||||
|
||||
print("=== mindnet Retriever Smoke-Test ===")
|
||||
print(f"URL: {args.url}")
|
||||
print(f"Mode: {args.mode}")
|
||||
print(f"Query: {args.query!r}")
|
||||
print(f"Top-K: {args.top_k}")
|
||||
print()
|
||||
|
||||
results = data.get("results") or []
|
||||
print(f"used_mode: {data.get('used_mode')}")
|
||||
print(f"latency: {data.get('latency_ms')} ms")
|
||||
print(f"Treffer: {len(results)}")
|
||||
print()
|
||||
|
||||
for i, hit in enumerate(results, start=1):
|
||||
nid = hit.get("note_id")
|
||||
cid = hit.get("node_id")
|
||||
tot = hit.get("total_score")
|
||||
sem = hit.get("semantic_score")
|
||||
edge = hit.get("edge_bonus")
|
||||
cen = hit.get("centrality_bonus")
|
||||
src = hit.get("source") or {}
|
||||
path = src.get("path")
|
||||
sec = src.get("section")
|
||||
|
||||
print(f"[{i}] note_id={nid} node_id={cid}")
|
||||
try:
|
||||
print(
|
||||
f" total={float(tot):.4f} "
|
||||
f"(semantic={float(sem):.4f} edge={float(edge):.4f} centrality={float(cen):.4f})"
|
||||
)
|
||||
except Exception:
|
||||
print(f" total={tot}, semantic={sem}, edge={edge}, centrality={cen}")
|
||||
|
||||
if path or sec:
|
||||
print(f" source: {path or '-'} :: {sec or '-'}")
|
||||
print()
|
||||
|
||||
return 0
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
raise SystemExit(main())
|
||||
@@ -0,0 +1,48 @@
|
||||
from app.core import retriever as r
|
||||
from app.models.dto import QueryRequest
|
||||
import app.core.qdrant as qdr
|
||||
import app.core.qdrant_points as qp
|
||||
import app.services.embeddings_client as ec
|
||||
|
||||
|
||||
class _DummyClient:
|
||||
pass
|
||||
|
||||
|
||||
def _fake_get_client(cfg):
|
||||
return _DummyClient()
|
||||
|
||||
|
||||
def _fake_embed_text(text: str):
|
||||
# Dimension egal, da wir search_chunks_by_vector faken
|
||||
return [0.0] * 384
|
||||
|
||||
|
||||
def _fake_search_chunks_by_vector_weights(client, prefix, vector, top=10, filters=None):
|
||||
# Zwei Treffer mit unterschiedlichem retriever_weight
|
||||
return [
|
||||
("chunk:low", 0.9, {"note_id": "note:low", "path": "low.md", "section_title": "S1", "retriever_weight": 0.2}),
|
||||
("chunk:high", 0.8, {"note_id": "note:high", "path": "high.md", "section_title": "S2", "retriever_weight": 1.5}),
|
||||
]
|
||||
|
||||
|
||||
def test_retriever_weight_influences_ranking(monkeypatch):
|
||||
# Konfiguration: Qdrant & Embedding faken
|
||||
monkeypatch.setattr(qdr, "get_client", _fake_get_client)
|
||||
monkeypatch.setattr(ec, "embed_text", _fake_embed_text)
|
||||
monkeypatch.setattr(qp, "search_chunks_by_vector", _fake_search_chunks_by_vector_weights)
|
||||
|
||||
req = QueryRequest(mode="semantic", query="karate", top_k=2)
|
||||
resp = r.semantic_retrieve(req)
|
||||
|
||||
assert len(resp.results) == 2
|
||||
|
||||
first = resp.results[0]
|
||||
second = resp.results[1]
|
||||
|
||||
# Semantic Scores: 0.9 (low), 0.8 (high)
|
||||
# Total Scores: 0.9 * 0.2 = 0.18 vs. 0.8 * 1.5 = 1.2 -> high sollte oben liegen
|
||||
assert first.note_id == "note:high"
|
||||
assert second.note_id == "note:low"
|
||||
|
||||
assert first.total_score > second.total_score
|
||||
Reference in new issue
Block a user