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:
1 parent
cd5056d4c9
commit
d35bdc64b9
5 files changed
+202
-91
No files matched your search
@@ -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)
|
||||
prov_w = 1.0 if prov == "explicit" else (0.9 if prov == "smart" else 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)
|
||||
|
||||
# B. Intent Boost Multiplikator (Vom Router dynamisch injiziert)
|
||||
kind = data.get("kind")
|
||||
intent_multiplier = boost_edges.get(kind, 1.0)
|
||||
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])
|
||||
|
||||
# Finales Gewicht setzen (Basis * Provenance * Intent)
|
||||
data["weight"] = data.get("weight", 1.0) * prov_w * intent_multiplier
|
||||
# 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
|
||||
kind = e.get("kind")
|
||||
intent_multiplier = boost_edges.get(kind, 1.0)
|
||||
|
||||
# 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)
|
||||
Reference in new issue
Block a user