Implement WP-15c enhancements across graph and retrieval modules, including full metadata support for Super-Edge aggregation and Note-Level Diversity Pooling. Update scoring logic to reflect new edge handling and improve retrieval accuracy. Version updates to reflect these changes.
This commit is contained in:
parent
cd5056d4c9
commit
d35bdc64b9
|
|
@ -1,13 +1,15 @@
|
|||
"""
|
||||
FILE: app/core/graph/graph_db_adapter.py
|
||||
DESCRIPTION: Datenbeschaffung aus Qdrant für den Graphen.
|
||||
AUDIT v1.1.0: Nutzt nun die zentrale database-Infrastruktur für Namen.
|
||||
AUDIT v1.1.1: Volle Unterstützung für WP-15c Metadaten.
|
||||
Stellt sicher, dass 'target_section' und 'provenance' für die
|
||||
Super-Edge-Aggregation im Retriever geladen werden.
|
||||
"""
|
||||
from typing import List, Dict, Optional
|
||||
from qdrant_client import QdrantClient
|
||||
from qdrant_client.http import models as rest
|
||||
|
||||
# ENTSCHEIDENDER FIX: Nutzt die neue Infrastruktur für konsistente Collection-Namen
|
||||
# Nutzt die zentrale Infrastruktur für konsistente Collection-Namen (WP-14)
|
||||
from app.core.database import collection_names
|
||||
|
||||
def fetch_edges_from_qdrant(
|
||||
|
|
@ -19,14 +21,16 @@ def fetch_edges_from_qdrant(
|
|||
) -> List[Dict]:
|
||||
"""
|
||||
Holt Edges aus der Datenbank basierend auf Seed-IDs.
|
||||
Filtert auf source_id, target_id oder note_id.
|
||||
WP-15c: Erhält alle Metadaten für das Note-Level Diversity Pooling.
|
||||
"""
|
||||
if not seeds or limit <= 0:
|
||||
return []
|
||||
|
||||
# Konsistente Namensauflösung via database-Paket
|
||||
# Rückgabe: (notes_col, chunks_col, edges_col)
|
||||
_, _, edges_col = collection_names(prefix)
|
||||
|
||||
# Wir suchen Kanten, bei denen die Seed-IDs entweder Quelle, Ziel oder Kontext-Note sind.
|
||||
seed_conditions = []
|
||||
for field in ("source_id", "target_id", "note_id"):
|
||||
for s in seeds:
|
||||
|
|
@ -35,6 +39,7 @@ def fetch_edges_from_qdrant(
|
|||
)
|
||||
seeds_filter = rest.Filter(should=seed_conditions) if seed_conditions else None
|
||||
|
||||
# Optionaler Filter auf spezifische Kanten-Typen (z.B. für Intent-Routing)
|
||||
type_filter = None
|
||||
if edge_types:
|
||||
type_conds = [
|
||||
|
|
@ -52,6 +57,7 @@ def fetch_edges_from_qdrant(
|
|||
flt = rest.Filter(must=must) if must else None
|
||||
|
||||
# Abfrage via Qdrant Scroll API
|
||||
# WICHTIG: with_payload=True lädt alle Metadaten (target_section, provenance etc.)
|
||||
pts, _ = client.scroll(
|
||||
collection_name=edges_col,
|
||||
scroll_filter=flt,
|
||||
|
|
@ -60,4 +66,6 @@ def fetch_edges_from_qdrant(
|
|||
with_vectors=False,
|
||||
)
|
||||
|
||||
# Wir geben das vollständige Payload zurück, damit der Retriever
|
||||
# alle Signale für die Super-Edge-Aggregation und das Scoring hat.
|
||||
return [dict(p.payload) for p in pts if p.payload]
|
||||
|
|
@ -1,10 +1,10 @@
|
|||
"""
|
||||
FILE: app/core/graph/graph_derive_edges.py
|
||||
DESCRIPTION: Hauptlogik zur Kanten-Aggregation und De-Duplizierung.
|
||||
AUDIT:
|
||||
- Nutzt parse_link_target
|
||||
- Übergibt Section als 'variant' an ID-Gen
|
||||
- Dedup basiert jetzt auf Edge-ID (erlaubt Multigraph für Sections)
|
||||
WP-15b/c Audit:
|
||||
- Präzises Sektions-Splitting via parse_link_target.
|
||||
- Eindeutige ID-Generierung pro Sektions-Variante (Multigraph).
|
||||
- Ermöglicht dem Retriever die Super-Edge-Aggregation.
|
||||
"""
|
||||
from typing import List, Optional, Dict, Tuple
|
||||
from .graph_utils import (
|
||||
|
|
@ -21,31 +21,45 @@ def build_edges_for_note(
|
|||
note_level_references: Optional[List[str]] = None,
|
||||
include_note_scope_refs: bool = False,
|
||||
) -> List[dict]:
|
||||
"""Erzeugt und aggregiert alle Kanten für eine Note (WP-15b)."""
|
||||
"""
|
||||
Erzeugt und aggregiert alle Kanten für eine Note.
|
||||
Sorgt für die physische Trennung von Sektions-Links via Edge-ID.
|
||||
"""
|
||||
edges: List[dict] = []
|
||||
# note_type für die Ermittlung der edge_defaults (types.yaml)
|
||||
note_type = _get(chunks[0], "type") if chunks else "concept"
|
||||
|
||||
# 1) Struktur-Kanten (belongs_to, next/prev)
|
||||
# 1) Struktur-Kanten (Internal: belongs_to, next/prev)
|
||||
# Diese erhalten die Provenienz 'structure' und sind in der Registry geschützt.
|
||||
for idx, ch in enumerate(chunks):
|
||||
cid = _get(ch, "chunk_id", "id")
|
||||
if not cid: continue
|
||||
|
||||
# Verbindung Chunk -> Note
|
||||
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": "structure", "rule_id": "structure:belongs_to", "confidence": PROVENANCE_PRIORITY["structure:belongs_to"]
|
||||
"chunk_id": cid,
|
||||
"edge_id": _mk_edge_id("belongs_to", cid, note_id, "chunk", "structure:belongs_to"),
|
||||
"provenance": "structure",
|
||||
"rule_id": "structure:belongs_to",
|
||||
"confidence": PROVENANCE_PRIORITY["structure:belongs_to"]
|
||||
}))
|
||||
|
||||
# Horizontale Verkettung (Ordnung)
|
||||
if idx < len(chunks) - 1:
|
||||
next_id = _get(chunks[idx+1], "chunk_id", "id")
|
||||
if next_id:
|
||||
edges.append(_edge("next", "chunk", cid, next_id, note_id, {
|
||||
"chunk_id": cid, "edge_id": _mk_edge_id("next", cid, next_id, "chunk", "structure:order"),
|
||||
"chunk_id": cid,
|
||||
"edge_id": _mk_edge_id("next", cid, next_id, "chunk", "structure:order"),
|
||||
"provenance": "structure", "rule_id": "structure:order", "confidence": PROVENANCE_PRIORITY["structure:order"]
|
||||
}))
|
||||
edges.append(_edge("prev", "chunk", next_id, cid, note_id, {
|
||||
"chunk_id": next_id, "edge_id": _mk_edge_id("prev", next_id, cid, "chunk", "structure:order"),
|
||||
"chunk_id": next_id,
|
||||
"edge_id": _mk_edge_id("prev", next_id, cid, "chunk", "structure:order"),
|
||||
"provenance": "structure", "rule_id": "structure:order", "confidence": PROVENANCE_PRIORITY["structure:order"]
|
||||
}))
|
||||
|
||||
# 2) Inhaltliche Kanten
|
||||
# 2) Inhaltliche Kanten (Explicit & Candidate Pool)
|
||||
reg = load_types_registry()
|
||||
defaults = get_edge_defaults_for(note_type, reg)
|
||||
refs_all: List[str] = []
|
||||
|
|
@ -55,7 +69,7 @@ def build_edges_for_note(
|
|||
if not cid: continue
|
||||
raw = _get(ch, "window") or _get(ch, "text") or ""
|
||||
|
||||
# Typed & Candidate Pool (WP-15b Integration)
|
||||
# A. Typed Relations (Inline [[rel:kind|target]])
|
||||
typed, rem = extract_typed_relations(raw)
|
||||
for k, raw_t in typed:
|
||||
t, sec = parse_link_target(raw_t, note_id)
|
||||
|
|
@ -63,14 +77,14 @@ def build_edges_for_note(
|
|||
|
||||
payload = {
|
||||
"chunk_id": cid,
|
||||
# Variant=sec sorgt für eindeutige ID pro Abschnitt
|
||||
# WP-Fix: Variant=sec sorgt für eindeutige ID pro Sektion
|
||||
"edge_id": _mk_edge_id(k, cid, t, "chunk", "inline:rel", variant=sec),
|
||||
"provenance": "explicit", "rule_id": "inline:rel", "confidence": PROVENANCE_PRIORITY["inline:rel"]
|
||||
}
|
||||
if sec: payload["target_section"] = sec
|
||||
|
||||
edges.append(_edge(k, "chunk", cid, t, note_id, payload))
|
||||
|
||||
# B. Candidate Pool (WP-15b Validierte KI-Kanten)
|
||||
pool = ch.get("candidate_pool") or ch.get("candidate_edges") or []
|
||||
for cand in pool:
|
||||
raw_t, k, p = cand.get("to"), cand.get("kind", "related_to"), cand.get("provenance", "semantic_ai")
|
||||
|
|
@ -82,10 +96,9 @@ def build_edges_for_note(
|
|||
"provenance": p, "rule_id": f"candidate:{p}", "confidence": PROVENANCE_PRIORITY.get(p, 0.90)
|
||||
}
|
||||
if sec: payload["target_section"] = sec
|
||||
|
||||
edges.append(_edge(k, "chunk", cid, t, note_id, payload))
|
||||
|
||||
# Callouts & Wikilinks
|
||||
# C. Callouts (> [!edge])
|
||||
call_pairs, rem2 = extract_callout_relations(rem)
|
||||
for k, raw_t in call_pairs:
|
||||
t, sec = parse_link_target(raw_t, note_id)
|
||||
|
|
@ -97,9 +110,9 @@ def build_edges_for_note(
|
|||
"provenance": "explicit", "rule_id": "callout:edge", "confidence": PROVENANCE_PRIORITY["callout:edge"]
|
||||
}
|
||||
if sec: payload["target_section"] = sec
|
||||
|
||||
edges.append(_edge(k, "chunk", cid, t, note_id, payload))
|
||||
|
||||
# D. Standard Wikilinks & Typ-Defaults
|
||||
refs = extract_wikilinks(rem2)
|
||||
for raw_r in refs:
|
||||
r, sec = parse_link_target(raw_r, note_id)
|
||||
|
|
@ -111,9 +124,9 @@ def build_edges_for_note(
|
|||
"provenance": "explicit", "rule_id": "explicit:wikilink", "confidence": PROVENANCE_PRIORITY["explicit:wikilink"]
|
||||
}
|
||||
if sec: payload["target_section"] = sec
|
||||
|
||||
edges.append(_edge("references", "chunk", cid, r, note_id, payload))
|
||||
|
||||
# Automatische Kanten-Vererbung aus types.yaml
|
||||
for rel in defaults:
|
||||
if rel != "references":
|
||||
def_payload = {
|
||||
|
|
@ -124,13 +137,10 @@ def build_edges_for_note(
|
|||
if sec: def_payload["target_section"] = sec
|
||||
edges.append(_edge(rel, "chunk", cid, r, note_id, def_payload))
|
||||
|
||||
# Für Note-Scope Sammlung nutzen wir den Original-String zur Dedup, aber gesäubert
|
||||
refs_all.extend([parse_link_target(r, note_id)[0] for r in refs])
|
||||
|
||||
# 3) Note-Scope & De-Duplizierung
|
||||
# 3) Note-Scope (Grobe Struktur-Verbindungen)
|
||||
if include_note_scope_refs:
|
||||
# refs_all ist jetzt schon gesäubert (nur Targets)
|
||||
# note_level_references müssen auch gesäubert werden
|
||||
cleaned_note_refs = [parse_link_target(r, note_id)[0] for r in (note_level_references or [])]
|
||||
refs_note = _dedupe_seq((refs_all or []) + cleaned_note_refs)
|
||||
|
||||
|
|
@ -140,17 +150,19 @@ def build_edges_for_note(
|
|||
"edge_id": _mk_edge_id("references", note_id, r, "note", "explicit:note_scope"),
|
||||
"provenance": "explicit", "confidence": PROVENANCE_PRIORITY["explicit:note_scope"]
|
||||
}))
|
||||
# Backlinks zur Stärkung der Bidirektionalität
|
||||
edges.append(_edge("backlink", "note", r, note_id, note_id, {
|
||||
"edge_id": _mk_edge_id("backlink", r, note_id, "note", "derived:backlink"),
|
||||
"provenance": "rule", "confidence": PROVENANCE_PRIORITY["derived:backlink"]
|
||||
}))
|
||||
|
||||
# Deduplizierung: Wir nutzen jetzt die EDGE-ID als Schlüssel.
|
||||
# Da die Edge-ID nun 'variant' (Section) enthält, bleiben unterschiedliche Sections erhalten.
|
||||
# 4) De-Duplizierung (In-Place)
|
||||
# Da die EDGE-ID nun die Sektion (variant) enthält, bleiben Links auf
|
||||
# unterschiedliche Abschnitte derselben Note erhalten.
|
||||
unique_map: Dict[str, dict] = {}
|
||||
for e in edges:
|
||||
eid = e["edge_id"]
|
||||
# Bei Konflikt (gleiche ID = exakt gleiche Kante und Section) gewinnt die höhere Confidence
|
||||
# Höhere Confidence gewinnt bei identischer ID
|
||||
if eid not in unique_map or e.get("confidence", 0) > unique_map[eid].get("confidence", 0):
|
||||
unique_map[eid] = e
|
||||
|
||||
|
|
|
|||
|
|
@ -2,8 +2,9 @@
|
|||
FILE: app/core/graph/graph_subgraph.py
|
||||
DESCRIPTION: In-Memory Repräsentation eines Graphen für Scoring und Analyse.
|
||||
Zentrale Komponente für die Graph-Expansion (BFS) und Bonus-Berechnung.
|
||||
MODULARISIERUNG: Teil des graph-Pakets (WP-14).
|
||||
VERSION: 1.1.0
|
||||
WP-15c Update: Erhalt von Metadaten (target_section, provenance)
|
||||
für präzises Retrieval-Reasoning.
|
||||
VERSION: 1.2.0
|
||||
STATUS: Active
|
||||
"""
|
||||
import math
|
||||
|
|
@ -22,6 +23,7 @@ class Subgraph:
|
|||
"""
|
||||
|
||||
def __init__(self) -> None:
|
||||
# adj speichert nun vollständige Payloads statt nur Tripel
|
||||
self.adj: DefaultDict[str, List[Dict]] = defaultdict(list)
|
||||
self.reverse_adj: DefaultDict[str, List[Dict]] = defaultdict(list)
|
||||
self.in_degree: DefaultDict[str, int] = defaultdict(int)
|
||||
|
|
@ -30,31 +32,49 @@ class Subgraph:
|
|||
def add_edge(self, e: Dict) -> None:
|
||||
"""
|
||||
Fügt eine Kante hinzu und aktualisiert Indizes.
|
||||
Unterstützt Kontext-Notes für verbesserte Graph-Konnektivität.
|
||||
WP-15c: Speichert das vollständige Payload für den Explanation Layer.
|
||||
"""
|
||||
src = e.get("source")
|
||||
tgt = e.get("target")
|
||||
kind = e.get("kind")
|
||||
weight = e.get("weight", EDGE_BASE_WEIGHTS.get(kind, 0.0))
|
||||
|
||||
# Das gesamte Payload wird als Kanten-Objekt behalten
|
||||
# Wir stellen sicher, dass alle relevanten Metadaten vorhanden sind
|
||||
edge_data = {
|
||||
"source": src,
|
||||
"target": tgt,
|
||||
"kind": kind,
|
||||
"weight": e.get("weight", EDGE_BASE_WEIGHTS.get(kind, 0.0)),
|
||||
"provenance": e.get("provenance", "rule"),
|
||||
"confidence": e.get("confidence", 1.0),
|
||||
"target_section": e.get("target_section"), # Essentiell für Präzision
|
||||
"is_super_edge": e.get("is_super_edge", False)
|
||||
}
|
||||
|
||||
owner = e.get("note_id")
|
||||
|
||||
if not src or not tgt:
|
||||
return
|
||||
|
||||
# 1. Forward-Kante
|
||||
self.adj[src].append({"target": tgt, "kind": kind, "weight": weight})
|
||||
self.adj[src].append(edge_data)
|
||||
self.out_degree[src] += 1
|
||||
self.in_degree[tgt] += 1
|
||||
|
||||
# 2. Reverse-Kante (für WP-04b Explanation Layer)
|
||||
self.reverse_adj[tgt].append({"source": src, "kind": kind, "weight": weight})
|
||||
# 2. Reverse-Kante (für Explanation Layer & Backlinks)
|
||||
self.reverse_adj[tgt].append(edge_data)
|
||||
|
||||
# 3. Kontext-Note Handling (erhöht die Zentralität der Parent-Note)
|
||||
if owner and owner != src:
|
||||
self.adj[owner].append({"target": tgt, "kind": kind, "weight": weight})
|
||||
# Wir erstellen eine virtuelle Kontext-Kante
|
||||
ctx_edge = edge_data.copy()
|
||||
ctx_edge["source"] = owner
|
||||
ctx_edge["via_context"] = True
|
||||
|
||||
self.adj[owner].append(ctx_edge)
|
||||
self.out_degree[owner] += 1
|
||||
if owner != tgt:
|
||||
self.reverse_adj[tgt].append({"source": owner, "kind": kind, "weight": weight, "via_context": True})
|
||||
self.reverse_adj[tgt].append(ctx_edge)
|
||||
self.in_degree[owner] += 1
|
||||
|
||||
def aggregate_edge_bonus(self, node_id: str) -> float:
|
||||
|
|
@ -73,14 +93,15 @@ class Subgraph:
|
|||
indeg = self.in_degree.get(node_id, 0)
|
||||
if indeg <= 0:
|
||||
return 0.0
|
||||
# math.log1p(x) entspricht log(1+x)
|
||||
return min(math.log1p(indeg) / 10.0, 0.15)
|
||||
|
||||
def get_outgoing_edges(self, node_id: str) -> List[Dict[str, Any]]:
|
||||
"""Gibt alle ausgehenden Kanten einer Node zurück."""
|
||||
"""Gibt alle ausgehenden Kanten einer Node inkl. Metadaten zurück."""
|
||||
return self.adj.get(node_id, [])
|
||||
|
||||
def get_incoming_edges(self, node_id: str) -> List[Dict[str, Any]]:
|
||||
"""Gibt alle eingehenden Kanten einer Node zurück."""
|
||||
"""Gibt alle eingehenden Kanten einer Node inkl. Metadaten zurück."""
|
||||
return self.reverse_adj.get(node_id, [])
|
||||
|
||||
|
||||
|
|
@ -111,13 +132,19 @@ def expand(
|
|||
src, tgt = pl.get("source_id"), pl.get("target_id")
|
||||
if not src or not tgt: continue
|
||||
|
||||
sg.add_edge({
|
||||
# WP-15c: Wir übergeben das vollständige Payload an add_edge
|
||||
edge_payload = {
|
||||
"source": src,
|
||||
"target": tgt,
|
||||
"kind": pl.get("kind", "edge"),
|
||||
"weight": calculate_edge_weight(pl),
|
||||
"note_id": pl.get("note_id"),
|
||||
})
|
||||
"provenance": pl.get("provenance", "rule"),
|
||||
"confidence": pl.get("confidence", 1.0),
|
||||
"target_section": pl.get("target_section")
|
||||
}
|
||||
|
||||
sg.add_edge(edge_payload)
|
||||
|
||||
# BFS Logik: Neue Ziele in die nächste Frontier aufnehmen
|
||||
if tgt not in visited:
|
||||
|
|
|
|||
|
|
@ -1,9 +1,8 @@
|
|||
"""
|
||||
FILE: app/core/retrieval/retriever.py
|
||||
DESCRIPTION: Haupt-Schnittstelle für die Suche. Orchestriert Vektorsuche und Graph-Expansion.
|
||||
Nutzt retriever_scoring.py für die WP-22 Logik.
|
||||
MODULARISIERUNG: Verschoben in das retrieval-Paket für WP-14.
|
||||
VERSION: 0.6.16
|
||||
WP-15c Update: Note-Level Diversity Pooling & Super-Edge Aggregation.
|
||||
VERSION: 0.7.0
|
||||
STATUS: Active
|
||||
DEPENDENCIES: app.config, app.models.dto, app.core.database*, app.core.graph_adapter
|
||||
"""
|
||||
|
|
@ -13,6 +12,7 @@ import os
|
|||
import time
|
||||
import logging
|
||||
from typing import Any, Dict, List, Tuple, Iterable, Optional
|
||||
from collections import defaultdict
|
||||
|
||||
from app.config import get_settings
|
||||
from app.models.dto import (
|
||||
|
|
@ -89,7 +89,6 @@ def _build_explanation(
|
|||
) -> Explanation:
|
||||
"""
|
||||
Transformiert mathematische Scores und Graph-Signale in eine menschenlesbare Erklärung.
|
||||
Behebt Pydantic ValidationErrors durch explizite String-Sicherung.
|
||||
"""
|
||||
_, edge_w_cfg, _ = get_weights()
|
||||
base_val = scoring_debug["base_val"]
|
||||
|
|
@ -116,12 +115,22 @@ def _build_explanation(
|
|||
elif semantic_score > 0.70:
|
||||
reasons.append(Reason(kind="semantic", message="Inhaltliche Übereinstimmung.", score_impact=base_val))
|
||||
|
||||
# 3. Gründe für Typ und Lifecycle
|
||||
# 3. Gründe für Typ und Lifecycle (WP-25 Vorbereitung)
|
||||
type_weight = float(payload.get("retriever_weight", 1.0))
|
||||
if type_weight != 1.0:
|
||||
msg = "Bevorzugt" if type_weight > 1.0 else "De-priorisiert"
|
||||
reasons.append(Reason(kind="type", message=f"{msg} durch Typ-Profil.", score_impact=base_val * (type_weight - 1.0)))
|
||||
|
||||
# NEU: Explizite Ausweisung des Lifecycle-Status (WP-22)
|
||||
status_mult = scoring_debug.get("status_multiplier", 1.0)
|
||||
if status_mult != 1.0:
|
||||
status_msg = "Belohnt (Stable)" if status_mult > 1.0 else "De-priorisiert (Draft)"
|
||||
reasons.append(Reason(
|
||||
kind="status",
|
||||
message=f"{status_msg} durch Content-Lifecycle.",
|
||||
score_impact=semantic_score * (status_mult - 1.0)
|
||||
))
|
||||
|
||||
# 4. Kanten-Verarbeitung (Graph-Intelligence)
|
||||
if subgraph and target_note_id and scoring_debug["edge_bonus"] > 0:
|
||||
raw_edges = []
|
||||
|
|
@ -131,7 +140,6 @@ def _build_explanation(
|
|||
raw_edges.extend(subgraph.get_outgoing_edges(target_note_id) or [])
|
||||
|
||||
for edge in raw_edges:
|
||||
# FIX: Zwingende String-Konvertierung für Pydantic-Stabilität
|
||||
src = str(edge.get("source") or "note_root")
|
||||
tgt = str(edge.get("target") or target_note_id or "unknown_target")
|
||||
kind = str(edge.get("kind", "related_to"))
|
||||
|
|
@ -187,10 +195,14 @@ def _build_hits_from_semantic(
|
|||
explain: bool = False,
|
||||
dynamic_edge_boosts: Dict[str, float] = None
|
||||
) -> QueryResponse:
|
||||
"""Wandelt semantische Roh-Treffer in bewertete QueryHits um."""
|
||||
"""
|
||||
Wandelt semantische Roh-Treffer in bewertete QueryHits um.
|
||||
WP-15c: Implementiert Note-Level Diversity Pooling.
|
||||
"""
|
||||
t0 = time.time()
|
||||
enriched = []
|
||||
|
||||
# Erstes Scoring für alle Kandidaten
|
||||
for pid, semantic_score, payload in hits:
|
||||
edge_bonus, cent_bonus = 0.0, 0.0
|
||||
target_id = payload.get("note_id")
|
||||
|
|
@ -202,15 +214,30 @@ def _build_hits_from_semantic(
|
|||
except Exception:
|
||||
pass
|
||||
|
||||
# Mathematisches Scoring via WP-22 Engine
|
||||
debug_data = compute_wp22_score(
|
||||
semantic_score, payload, edge_bonus, cent_bonus, dynamic_edge_boosts
|
||||
)
|
||||
enriched.append((pid, semantic_score, payload, debug_data))
|
||||
|
||||
# Sortierung nach finalem mathematischen Score
|
||||
# 1. Sortierung nach finalem mathematischen Score
|
||||
enriched_sorted = sorted(enriched, key=lambda h: h[3]["total"], reverse=True)
|
||||
limited_hits = enriched_sorted[: max(1, top_k)]
|
||||
|
||||
# 2. WP-15c: Note-Level Diversity Pooling
|
||||
# Wir behalten pro note_id nur den Hit mit dem höchsten total_score.
|
||||
# Dies verhindert, dass 10 Chunks derselben Note andere KeyNotes verdrängen.
|
||||
unique_note_hits = []
|
||||
seen_notes = set()
|
||||
|
||||
for item in enriched_sorted:
|
||||
_, _, payload, _ = item
|
||||
note_id = str(payload.get("note_id", "unknown"))
|
||||
|
||||
if note_id not in seen_notes:
|
||||
unique_note_hits.append(item)
|
||||
seen_notes.add(note_id)
|
||||
|
||||
# 3. Begrenzung auf top_k nach dem Diversity-Pooling
|
||||
limited_hits = unique_note_hits[: max(1, top_k)]
|
||||
|
||||
results: List[QueryHit] = []
|
||||
for pid, s_score, pl, dbg in limited_hits:
|
||||
|
|
@ -225,7 +252,6 @@ def _build_hits_from_semantic(
|
|||
applied_boosts=dynamic_edge_boosts
|
||||
)
|
||||
|
||||
# Payload Text-Feld normalisieren
|
||||
text_content = pl.get("page_content") or pl.get("text") or pl.get("content", "[Kein Text]")
|
||||
|
||||
results.append(QueryHit(
|
||||
|
|
@ -250,14 +276,14 @@ def _build_hits_from_semantic(
|
|||
def hybrid_retrieve(req: QueryRequest) -> QueryResponse:
|
||||
"""
|
||||
Die Haupt-Einstiegsfunktion für die hybride Suche.
|
||||
Kombiniert Vektorsuche mit Graph-Expansion und WP-22 Gewichtung.
|
||||
WP-15c: Implementiert Edge-Aggregation (Super-Kanten).
|
||||
"""
|
||||
client, prefix = _get_client_and_prefix()
|
||||
vector = list(req.query_vector) if req.query_vector else _get_query_vector(req)
|
||||
top_k = req.top_k or 10
|
||||
|
||||
# 1. Semantische Seed-Suche
|
||||
hits = _semantic_hits(client, prefix, vector, top_k=top_k, filters=req.filters)
|
||||
# 1. Semantische Seed-Suche (Wir laden etwas mehr für das Pooling)
|
||||
hits = _semantic_hits(client, prefix, vector, top_k=top_k * 3, filters=req.filters)
|
||||
|
||||
# 2. Graph Expansion Konfiguration
|
||||
expand_cfg = req.expand if isinstance(req.expand, dict) else {}
|
||||
|
|
@ -266,39 +292,76 @@ def hybrid_retrieve(req: QueryRequest) -> QueryResponse:
|
|||
|
||||
subgraph: ga.Subgraph | None = None
|
||||
if depth > 0 and hits:
|
||||
# Start-IDs für den Graph-Traversal sammeln
|
||||
seed_ids = list({h[2].get("note_id") for h in hits if h[2].get("note_id")})
|
||||
|
||||
if seed_ids:
|
||||
try:
|
||||
# Subgraph aus RAM/DB laden
|
||||
subgraph = ga.expand(client, prefix, seed_ids, depth=depth, edge_types=expand_cfg.get("edge_types"))
|
||||
|
||||
# --- WP-22: Kanten-Gewichtung im RAM-Graphen vor Bonus-Berechnung ---
|
||||
if subgraph and hasattr(subgraph, "graph"):
|
||||
for _, _, data in subgraph.graph.edges(data=True):
|
||||
# A. Provenance Weighting (WP-22 Bonus für Herkunft)
|
||||
prov = data.get("provenance", "rule")
|
||||
# Belohnung: Explizite Links (1.0) > Smart (0.9) > Rule (0.7)
|
||||
# --- WP-15c: Edge-Aggregation & Deduplizierung (Super-Kanten) ---
|
||||
# Verhindert Score-Explosion durch multiple Links auf versch. Abschnitte.
|
||||
# Logik: 1. Kante zählt voll, weitere dämpfen auf Faktor 0.1.
|
||||
if subgraph and hasattr(subgraph, "adj"):
|
||||
for src, edge_list in subgraph.adj.items():
|
||||
# Gruppiere Kanten nach Ziel-Note (Deduplizierung ID_A -> ID_B)
|
||||
by_target = defaultdict(list)
|
||||
for e in edge_list:
|
||||
by_target[e["target"]].append(e)
|
||||
|
||||
aggregated_list = []
|
||||
for tgt, edges in by_target.items():
|
||||
if len(edges) > 1:
|
||||
# Sortiere: Stärkste Kante zuerst
|
||||
sorted_edges = sorted(edges, key=lambda x: x.get("weight", 0.0), reverse=True)
|
||||
primary = sorted_edges[0]
|
||||
|
||||
# Aggregiertes Gewicht berechnen (Sättigungs-Logik)
|
||||
total_w = primary.get("weight", 0.0)
|
||||
for secondary in sorted_edges[1:]:
|
||||
total_w += secondary.get("weight", 0.0) * 0.1
|
||||
|
||||
primary["weight"] = total_w
|
||||
primary["is_super_edge"] = True # Flag für Explanation Layer
|
||||
primary["edge_count"] = len(edges)
|
||||
aggregated_list.append(primary)
|
||||
else:
|
||||
aggregated_list.append(edges[0])
|
||||
|
||||
# In-Place Update der Adjazenzliste des Graphen
|
||||
subgraph.adj[src] = aggregated_list
|
||||
|
||||
# Re-Sync der In-Degrees für Centrality-Bonus (Aggregation konsistent halten)
|
||||
subgraph.in_degree = defaultdict(int)
|
||||
for src, edges in subgraph.adj.items():
|
||||
for e in edges:
|
||||
subgraph.in_degree[e["target"]] += 1
|
||||
|
||||
# --- WP-22: Kanten-Gewichtung (Provenance & Intent Boost) ---
|
||||
if subgraph and hasattr(subgraph, "adj"):
|
||||
for src, edges in subgraph.adj.items():
|
||||
for e in edges:
|
||||
# A. Provenance Weighting
|
||||
prov = e.get("provenance", "rule")
|
||||
prov_w = 1.0 if prov == "explicit" else (0.9 if prov == "smart" else 0.7)
|
||||
|
||||
# B. Intent Boost Multiplikator (Vom Router dynamisch injiziert)
|
||||
kind = data.get("kind")
|
||||
# B. Intent Boost Multiplikator
|
||||
kind = e.get("kind")
|
||||
intent_multiplier = boost_edges.get(kind, 1.0)
|
||||
|
||||
# Finales Gewicht setzen (Basis * Provenance * Intent)
|
||||
data["weight"] = data.get("weight", 1.0) * prov_w * intent_multiplier
|
||||
# Gewichtung anpassen
|
||||
e["weight"] = e.get("weight", 1.0) * prov_w * intent_multiplier
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"Graph Expansion failed: {e}")
|
||||
subgraph = None
|
||||
|
||||
# 3. Scoring & Explanation Generierung
|
||||
# top_k wird erst hier final angewandt
|
||||
return _build_hits_from_semantic(hits, top_k, "hybrid", subgraph, req.explain, boost_edges)
|
||||
|
||||
|
||||
def semantic_retrieve(req: QueryRequest) -> QueryResponse:
|
||||
"""Standard Vektorsuche ohne Graph-Einfluss (WP-02 Fallback)."""
|
||||
"""Standard Vektorsuche ohne Graph-Einfluss."""
|
||||
client, prefix = _get_client_and_prefix()
|
||||
vector = _get_query_vector(req)
|
||||
hits = _semantic_hits(client, prefix, vector, req.top_k or 10, req.filters)
|
||||
|
|
@ -308,5 +371,4 @@ def semantic_retrieve(req: QueryRequest) -> QueryResponse:
|
|||
class Retriever:
|
||||
"""Schnittstelle für die asynchrone Suche."""
|
||||
async def search(self, request: QueryRequest) -> QueryResponse:
|
||||
"""Führt eine hybride Suche aus."""
|
||||
return hybrid_retrieve(request)
|
||||
|
|
@ -1,11 +1,10 @@
|
|||
"""
|
||||
FILE: app/core/retrieval/retriever_scoring.py
|
||||
DESCRIPTION: Mathematische Kern-Logik für das WP-22 Scoring.
|
||||
DESCRIPTION: Mathematische Kern-Logik für das WP-22/WP-15c Scoring.
|
||||
Berechnet Relevanz-Scores basierend auf Semantik, Graph-Intelligence und Content Lifecycle.
|
||||
MODULARISIERUNG: Verschoben in das retrieval-Paket für WP-14.
|
||||
VERSION: 1.0.2
|
||||
FIX v1.0.3: Optimierte Interaktion zwischen Typ-Boost und Status-Dämpfung.
|
||||
VERSION: 1.0.3
|
||||
STATUS: Active
|
||||
DEPENDENCIES: app.config, typing
|
||||
"""
|
||||
import os
|
||||
import logging
|
||||
|
|
@ -23,10 +22,6 @@ logger = logging.getLogger(__name__)
|
|||
def get_weights() -> Tuple[float, float, float]:
|
||||
"""
|
||||
Liefert die Basis-Gewichtung (semantic, edge, centrality) aus der Konfiguration.
|
||||
Priorität:
|
||||
1. config/retriever.yaml (Scoring-Sektion)
|
||||
2. Umgebungsvariablen (RETRIEVER_W_*)
|
||||
3. System-Defaults (1.0, 0.0, 0.0)
|
||||
"""
|
||||
from app.config import get_settings
|
||||
settings = get_settings()
|
||||
|
|
@ -58,7 +53,7 @@ def get_status_multiplier(payload: Dict[str, Any]) -> float:
|
|||
|
||||
- stable: 1.2 (Belohnung für verifiziertes Wissen)
|
||||
- active: 1.0 (Standard-Gewichtung)
|
||||
- draft: 0.5 (Bestrafung für unfertige Fragmente)
|
||||
- draft: 0.5 (Dämpfung für unfertige Fragmente)
|
||||
"""
|
||||
status = str(payload.get("status", "active")).lower().strip()
|
||||
if status == "stable":
|
||||
|
|
@ -75,35 +70,42 @@ def compute_wp22_score(
|
|||
dynamic_edge_boosts: Optional[Dict[str, float]] = None
|
||||
) -> Dict[str, Any]:
|
||||
"""
|
||||
Die zentrale mathematische Scoring-Formel der Mindnet Intelligence.
|
||||
Implementiert das WP-22 Hybrid-Scoring (Semantic * Lifecycle * Graph).
|
||||
Die zentrale mathematische Scoring-Formel (WP-15c optimiert).
|
||||
Implementiert das Hybrid-Scoring (Semantic * Lifecycle * Graph).
|
||||
|
||||
FORMEL:
|
||||
Score = (Similarity * StatusMult) * (1 + (TypeWeight - 1) + ((EdgeW * EB + CentW * CB) * IntentBoost))
|
||||
LOGIK:
|
||||
1. Base = Similarity * StatusMult (Lifecycle-Filter).
|
||||
2. Boosts = (TypeBoost - 1) + (GraphBoni * IntentFactor).
|
||||
3. Final = Base * (1 + Boosts).
|
||||
|
||||
Returns:
|
||||
Dict mit dem finalen 'total' Score und allen mathematischen Zwischenwerten für den Explanation Layer.
|
||||
Der edge_bonus_raw enthält bereits die Super-Edge-Aggregation (WP-15c).
|
||||
"""
|
||||
sem_w, edge_w_cfg, cent_w_cfg = get_weights()
|
||||
status_mult = get_status_multiplier(payload)
|
||||
|
||||
# Retriever Weight (Type Boost aus types.yaml, z.B. 1.1 für Decisions)
|
||||
# Retriever Weight (Typ-Boost aus types.yaml, z.B. 1.1 für Decisions)
|
||||
node_weight = float(payload.get("retriever_weight", 1.0))
|
||||
|
||||
# 1. Berechnung des Base Scores (Semantik gewichtet durch Lifecycle-Status)
|
||||
# WICHTIG: Der Status wirkt hier als Multiplikator auf die Basis-Relevanz.
|
||||
base_val = float(semantic_score) * status_mult
|
||||
|
||||
# 2. Graph Boost Factor (Teil C: Intent-spezifische Verstärkung)
|
||||
# 2. Graph Boost Factor (Intent-spezifische Verstärkung aus decision_engine.yaml)
|
||||
# Erhöht das Gewicht des gesamten Graphen um 50%, wenn ein spezifischer Intent vorliegt.
|
||||
graph_boost_factor = 1.5 if dynamic_edge_boosts and (edge_bonus_raw > 0 or cent_bonus_raw > 0) else 1.0
|
||||
|
||||
# 3. Einzelne Graph-Komponenten berechnen
|
||||
# WP-15c Hinweis: edge_bonus_raw ist durch den retriever.py bereits gedämpft/aggregiert.
|
||||
edge_impact_final = (edge_w_cfg * edge_bonus_raw) * graph_boost_factor
|
||||
cent_impact_final = (cent_w_cfg * cent_bonus_raw) * graph_boost_factor
|
||||
|
||||
# 4. Finales Zusammenführen (Merging)
|
||||
# (node_weight - 1.0) sorgt dafür, dass ein Gewicht von 1.0 keinen Einfluss hat (neutral).
|
||||
total = base_val * (1.0 + (node_weight - 1.0) + edge_impact_final + cent_impact_final)
|
||||
# (node_weight - 1.0) wandelt das Gewicht in einen relativen Bonus um (z.B. 1.2 -> +0.2).
|
||||
# Alle Boni werden addiert und wirken dann auf den base_val.
|
||||
type_impact = node_weight - 1.0
|
||||
total_boost = 1.0 + type_impact + edge_impact_final + cent_impact_final
|
||||
|
||||
total = base_val * total_boost
|
||||
|
||||
# Sicherstellen, dass der Score niemals 0 oder negativ ist (Floor)
|
||||
final_score = max(0.0001, float(total))
|
||||
|
|
@ -114,7 +116,7 @@ def compute_wp22_score(
|
|||
"cent_bonus": float(cent_bonus_raw),
|
||||
"status_multiplier": status_mult,
|
||||
"graph_boost_factor": graph_boost_factor,
|
||||
"type_impact": node_weight - 1.0,
|
||||
"type_impact": type_impact,
|
||||
"base_val": base_val,
|
||||
"edge_impact_final": edge_impact_final,
|
||||
"cent_impact_final": cent_impact_final
|
||||
|
|
|
|||
Loading…
Reference in New Issue
Block a user