mehrdimensionale matrix für Kanten
This commit is contained in:
parent
a1a58727fd
commit
b815f6235f
|
|
@ -1,7 +1,12 @@
|
|||
"""
|
||||
app/services/discovery.py
|
||||
Service für Link-Vorschläge und Knowledge-Discovery (WP-11).
|
||||
Optimiert: Deduplizierung pro Notiz & Footer-Fokus für kurze Texte.
|
||||
|
||||
Features:
|
||||
- Sliding Window Analyse für lange Texte.
|
||||
- Footer-Scan für Projekt-Referenzen.
|
||||
- 'Matrix-Logic' für intelligente Kanten-Typen (Experience -> Value = based_on).
|
||||
- Async & Nomic-Embeddings kompatibel.
|
||||
"""
|
||||
import logging
|
||||
import asyncio
|
||||
|
|
@ -23,33 +28,42 @@ class DiscoveryService:
|
|||
self.registry = self._load_type_registry()
|
||||
|
||||
async def analyze_draft(self, text: str, current_type: str) -> Dict[str, Any]:
|
||||
"""
|
||||
Analysiert den Text und liefert Vorschläge mit kontext-sensitiven Kanten-Typen.
|
||||
"""
|
||||
suggestions = []
|
||||
|
||||
# Fallback, falls keine spezielle Regel greift
|
||||
default_edge_type = self._get_default_edge_type(current_type)
|
||||
|
||||
# Tracking-Sets für Deduplizierung (Wir merken uns NOTE-IDs, nicht Chunk-IDs)
|
||||
# Tracking-Sets für Deduplizierung (Wir merken uns NOTE-IDs)
|
||||
seen_target_note_ids = set()
|
||||
|
||||
# ---------------------------------------------------------
|
||||
# 1. Exact Match: Titel/Aliases
|
||||
# ---------------------------------------------------------
|
||||
# Holt Titel, Aliases UND Typen aus dem Index
|
||||
known_entities = self._fetch_all_titles_and_aliases()
|
||||
found_entities = self._find_entities_in_text(text, known_entities)
|
||||
|
||||
for entity in found_entities:
|
||||
# Duplikate vermeiden
|
||||
if entity["id"] in seen_target_note_ids:
|
||||
continue
|
||||
seen_target_note_ids.add(entity["id"])
|
||||
|
||||
# INTELLIGENTE KANTEN-LOGIK (MATRIX)
|
||||
target_type = entity.get("type", "concept")
|
||||
smart_edge = self._resolve_edge_type(current_type, target_type)
|
||||
|
||||
suggestions.append({
|
||||
"type": "exact_match",
|
||||
"text_found": entity["match"],
|
||||
"target_title": entity["title"],
|
||||
"target_id": entity["id"],
|
||||
"suggested_edge_type": default_edge_type,
|
||||
"suggested_markdown": f"[[rel:{default_edge_type} {entity['title']}]]",
|
||||
"suggested_edge_type": smart_edge,
|
||||
"suggested_markdown": f"[[rel:{smart_edge} {entity['title']}]]",
|
||||
"confidence": 1.0,
|
||||
"reason": f"Exakter Treffer: '{entity['match']}'"
|
||||
"reason": f"Exakter Treffer: '{entity['match']}' ({target_type})"
|
||||
})
|
||||
|
||||
# ---------------------------------------------------------
|
||||
|
|
@ -64,33 +78,33 @@ class DiscoveryService:
|
|||
# Ergebnisse verarbeiten
|
||||
for hits in results_list:
|
||||
for hit in hits:
|
||||
# WICHTIG: Note ID aus Payload holen (Chunk ID ist hit.node_id)
|
||||
note_id = hit.payload.get("note_id")
|
||||
if not note_id: continue
|
||||
|
||||
# Fallback, falls Payload leer (sollte nicht passieren)
|
||||
if not note_id:
|
||||
continue
|
||||
|
||||
# 1. Check: Haben wir diese NOTIZ schon? (Egal welcher Chunk)
|
||||
# Deduplizierung (Notiz-Ebene)
|
||||
if note_id in seen_target_note_ids:
|
||||
continue
|
||||
|
||||
# 2. Score Check (Threshold)
|
||||
# Score Check (Threshold 0.50 für nomic-embed-text)
|
||||
if hit.total_score > 0.50:
|
||||
seen_target_note_ids.add(note_id) # Blockiere weitere Chunks dieser Notiz
|
||||
seen_target_note_ids.add(note_id)
|
||||
|
||||
target_title = hit.payload.get("title") or "Unbekannt"
|
||||
suggested_md = f"[[rel:{default_edge_type} {target_title}]]"
|
||||
|
||||
# INTELLIGENTE KANTEN-LOGIK (MATRIX)
|
||||
# Den Typ der gefundenen Notiz aus dem Payload lesen
|
||||
target_type = hit.payload.get("type", "concept")
|
||||
smart_edge = self._resolve_edge_type(current_type, target_type)
|
||||
|
||||
suggestions.append({
|
||||
"type": "semantic_match",
|
||||
"text_found": (hit.source.get("text") or "")[:60] + "...",
|
||||
"target_title": target_title,
|
||||
"target_id": note_id, # Wir verlinken auf die Notiz, nicht den Chunk
|
||||
"suggested_edge_type": default_edge_type,
|
||||
"suggested_markdown": suggested_md,
|
||||
"target_id": note_id,
|
||||
"suggested_edge_type": smart_edge,
|
||||
"suggested_markdown": f"[[rel:{smart_edge} {target_title}]]",
|
||||
"confidence": round(hit.total_score, 2),
|
||||
"reason": f"Semantisch ähnlich ({hit.total_score:.2f})"
|
||||
"reason": f"Semantisch ähnlich zu {target_type} ({hit.total_score:.2f})"
|
||||
})
|
||||
|
||||
# Sortieren nach Confidence
|
||||
|
|
@ -103,34 +117,63 @@ class DiscoveryService:
|
|||
"suggestions": suggestions[:10]
|
||||
}
|
||||
|
||||
# --- Optimierte Sliding Windows ---
|
||||
# ---------------------------------------------------------
|
||||
# Core Logic: Die Matrix
|
||||
# ---------------------------------------------------------
|
||||
|
||||
def _resolve_edge_type(self, source_type: str, target_type: str) -> str:
|
||||
"""
|
||||
Entscheidungsmatrix für komplexe Verbindungen.
|
||||
Definiert, wie Typ A auf Typ B verlinken sollte.
|
||||
"""
|
||||
st = source_type.lower()
|
||||
tt = target_type.lower()
|
||||
|
||||
# Regeln für 'experience' (Erfahrungen)
|
||||
if st == "experience":
|
||||
if tt == "value": return "based_on"
|
||||
if tt == "principle": return "derived_from"
|
||||
if tt == "trip": return "part_of"
|
||||
if tt == "lesson": return "learned"
|
||||
if tt == "project": return "related_to" # oder belongs_to
|
||||
|
||||
# Regeln für 'project'
|
||||
if st == "project":
|
||||
if tt == "decision": return "depends_on"
|
||||
if tt == "concept": return "uses"
|
||||
if tt == "person": return "managed_by"
|
||||
|
||||
# Regeln für 'decision' (ADR)
|
||||
if st == "decision":
|
||||
if tt == "principle": return "compliant_with"
|
||||
if tt == "requirement": return "addresses"
|
||||
|
||||
# Fallback: Standard aus der types.yaml für den Source-Typ
|
||||
return self._get_default_edge_type(st)
|
||||
|
||||
# ---------------------------------------------------------
|
||||
# Sliding Windows
|
||||
# ---------------------------------------------------------
|
||||
|
||||
def _generate_search_queries(self, text: str) -> List[str]:
|
||||
"""
|
||||
Erzeugt intelligente Fenster.
|
||||
Besonderheit: Erzwingt 'Footer-Scan' auch bei kurzen Texten,
|
||||
damit "Referenzen am Ende" nicht im Kontext untergehen.
|
||||
Erzeugt intelligente Fenster + Footer Scan.
|
||||
"""
|
||||
text_len = len(text)
|
||||
if not text: return []
|
||||
|
||||
queries = []
|
||||
|
||||
# A) Der gesamte Text (oder Anfang) für den groben Kontext
|
||||
# Bei sehr kurzen Texten ist das alles.
|
||||
# 1. Start / Gesamtkontext
|
||||
queries.append(text[:600])
|
||||
|
||||
# B) Der "Footer-Scan" (Das Ende)
|
||||
# Wenn der Text > 150 Zeichen ist, nehmen wir die letzten 200 Zeichen separat.
|
||||
# Grund: Oft steht dort "Gehört zu Projekt X".
|
||||
# Wenn wir das isolieren, ist der Vektor "Projekt X" sehr rein.
|
||||
# 2. Footer-Scan (Wichtig für "Projekt"-Referenzen am Ende)
|
||||
if text_len > 150:
|
||||
footer = text[-250:]
|
||||
# Nur hinzufügen, wenn es sich signifikant vom Start unterscheidet
|
||||
if footer not in queries:
|
||||
queries.append(footer)
|
||||
|
||||
# C) Sliding Window für lange Texte (> 800 Chars)
|
||||
# 3. Sliding Window für lange Texte
|
||||
if text_len > 800:
|
||||
window_size = 500
|
||||
step = 1500
|
||||
|
|
@ -142,7 +185,9 @@ class DiscoveryService:
|
|||
|
||||
return queries
|
||||
|
||||
# --- Standard Helper (Unverändert) ---
|
||||
# ---------------------------------------------------------
|
||||
# Standard Helpers
|
||||
# ---------------------------------------------------------
|
||||
|
||||
async def _get_semantic_suggestions_async(self, text: str):
|
||||
req = QueryRequest(query=text, top_k=5, explain=False)
|
||||
|
|
@ -174,12 +219,21 @@ class DiscoveryService:
|
|||
col = f"{self.prefix}_notes"
|
||||
try:
|
||||
while True:
|
||||
res, next_page = self.client.scroll(collection_name=col, limit=1000, offset=next_page, with_payload=True, with_vectors=False)
|
||||
res, next_page = self.client.scroll(
|
||||
collection_name=col, limit=1000, offset=next_page,
|
||||
with_payload=True, with_vectors=False
|
||||
)
|
||||
for point in res:
|
||||
pl = point.payload or {}
|
||||
aliases = pl.get("aliases") or []
|
||||
if isinstance(aliases, str): aliases = [aliases]
|
||||
notes.append({"id": pl.get("note_id"), "title": pl.get("title"), "aliases": aliases})
|
||||
|
||||
notes.append({
|
||||
"id": pl.get("note_id"),
|
||||
"title": pl.get("title"),
|
||||
"aliases": aliases,
|
||||
"type": pl.get("type", "concept") # WICHTIG: Typ laden für Matrix
|
||||
})
|
||||
if next_page is None: break
|
||||
except Exception: pass
|
||||
return notes
|
||||
|
|
@ -188,12 +242,14 @@ class DiscoveryService:
|
|||
found = []
|
||||
text_lower = text.lower()
|
||||
for entity in entities:
|
||||
# Title Check
|
||||
title = entity.get("title")
|
||||
if title and title.lower() in text_lower:
|
||||
found.append({"match": title, "title": title, "id": entity["id"]})
|
||||
found.append({"match": title, "title": title, "id": entity["id"], "type": entity["type"]})
|
||||
continue
|
||||
# Alias Check
|
||||
for alias in entity.get("aliases", []):
|
||||
if str(alias).lower() in text_lower:
|
||||
found.append({"match": alias, "title": title, "id": entity["id"]})
|
||||
found.append({"match": alias, "title": title, "id": entity["id"], "type": entity["type"]})
|
||||
break
|
||||
return found
|
||||
Loading…
Reference in New Issue
Block a user