neue version
This commit is contained in:
1 parent
2c3ee8efd6
commit
136c3bb43f
6 files changed
+138
-94
No files matched your search
@@ -4,7 +4,7 @@ DESCRIPTION: Haupt-Ingestion-Logik.
|
||||
FIX: Korrekte Priorisierung von Frontmatter für chunk_profile und retriever_weight.
|
||||
Lade Chunk-Config basierend auf dem effektiven Profil, nicht nur dem Notiz-Typ.
|
||||
WP-22: Integration von Content Lifecycle (Status) und Edge Registry.
|
||||
VERSION: 2.8.1 (WP-22 Lifecycle & Registry)
|
||||
VERSION: 2.8.5 (WP-22 Lifecycle & Registry)
|
||||
STATUS: Active
|
||||
DEPENDENCIES: app.core.parser, app.core.note_payload, app.core.chunker, app.core.derive_edges, app.core.qdrant*, app.services.embeddings_client, app.services.edge_registry
|
||||
EXTERNAL_CONFIG: config/types.yaml
|
||||
|
||||
+52
-51
@@ -2,7 +2,7 @@
|
||||
FILE: app/core/retriever.py
|
||||
DESCRIPTION: Implementiert die Hybrid-Suche (Vektor + Graph-Expansion) und das Scoring-Modell (Explainability).
|
||||
WP-22 Update: Dynamic Edge Boosting & Lifecycle Scoring.
|
||||
VERSION: 0.6.1 (WP-22 Dynamic Scoring)
|
||||
VERSION: 0.6.5 (WP-22 Scoring Formula)
|
||||
STATUS: Active
|
||||
DEPENDENCIES: app.config, app.models.dto, app.core.qdrant*, app.services.embeddings_client, app.core.graph_adapter
|
||||
LAST_ANALYSIS: 2025-12-18
|
||||
@@ -101,45 +101,48 @@ def _semantic_hits(
|
||||
# --- WP-22 Helper: Lifecycle Multipliers (Teil A) ---
|
||||
def _get_status_multiplier(payload: Dict[str, Any]) -> float:
|
||||
"""
|
||||
WP-22: stable (1.2), active (1.0), draft (0.5).
|
||||
WP-22: stable (1.2), active/default (1.0), draft (0.5).
|
||||
"""
|
||||
status = str(payload.get("status", "active")).lower()
|
||||
if status == "stable": return 1.2
|
||||
if status == "active": return 1.0
|
||||
if status == "draft": return 0.5
|
||||
return 1.0
|
||||
|
||||
# --- WP-22: Dynamic Scoring Formula (Teil C) ---
|
||||
def _compute_total_score(
|
||||
semantic_score: float,
|
||||
payload: Dict[str, Any],
|
||||
edge_bonus: float = 0.0,
|
||||
cent_bonus: float = 0.0,
|
||||
edge_bonus_raw: float = 0.0,
|
||||
cent_bonus_raw: float = 0.0,
|
||||
dynamic_edge_boosts: Dict[str, float] = None
|
||||
) -> Tuple[float, float, float]:
|
||||
"""
|
||||
Berechnet total_score nach WP-22 Scoring Formel.
|
||||
WP-22 Mathematische Logik:
|
||||
Score = BaseScore * (1 + ConfigWeight + DynamicBoost)
|
||||
|
||||
Hierbei gilt:
|
||||
- BaseScore: semantic_similarity * status_multiplier
|
||||
- ConfigWeight: retriever_weight (Type Boost)
|
||||
- DynamicBoost: (edge_weight * edge_bonus) + (centrality_weight * centrality_bonus)
|
||||
"""
|
||||
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()
|
||||
|
||||
# 1. Base Score (Semantik * Lifecycle)
|
||||
status_mult = _get_status_multiplier(payload)
|
||||
|
||||
# Dynamic Edge Boosting (Teil C)
|
||||
# Globaler Bonus falls Kanten-spezifische Boosts aktiv sind (z.B. WHY Frage)
|
||||
# Die kanten-spezifische Gewichtung passiert bereits im Subgraph in hybrid_retrieve.
|
||||
final_edge_score = edge_w * edge_bonus
|
||||
if dynamic_edge_boosts and edge_bonus > 0:
|
||||
# Globaler Verstärker für Graph-Signale bei spezifischen Intents
|
||||
final_edge_score *= 1.5
|
||||
|
||||
total = (sem_w * float(semantic_score) * weight * status_mult) + final_edge_score + (cent_w * cent_bonus)
|
||||
return float(total), float(edge_bonus), float(cent_bonus)
|
||||
base_score = float(semantic_score) * status_mult
|
||||
|
||||
# 2. Config Weight (Static Type Boost)
|
||||
config_weight = float(payload.get("retriever_weight", 1.0)) - 1.0 # 1.0 ist neutral
|
||||
|
||||
# 3. Dynamic Boost (Graph-Signale)
|
||||
_sem_w, edge_w_cfg, cent_w_cfg = _get_scoring_weights()
|
||||
dynamic_boost = (edge_w_cfg * edge_bonus_raw) + (cent_w_cfg * cent_bonus_raw)
|
||||
|
||||
# Falls Intent-Boosts vorliegen, verstärken wir den Dynamic Boost
|
||||
if dynamic_edge_boosts and (edge_bonus_raw > 0 or cent_bonus_raw > 0):
|
||||
dynamic_boost *= 1.5
|
||||
|
||||
total = base_score * (1.0 + config_weight + dynamic_boost)
|
||||
return float(total), float(edge_bonus_raw), float(cent_bonus_raw)
|
||||
|
||||
|
||||
# --- WP-04b Explanation Logic ---
|
||||
@@ -153,22 +156,21 @@ def _build_explanation(
|
||||
node_key: Optional[str]
|
||||
) -> Explanation:
|
||||
"""Erstellt ein Explanation-Objekt (WP-04b)."""
|
||||
sem_w, _edge_w, _cent_w = _get_scoring_weights()
|
||||
# Scoring weights erneut laden für Reason-Details
|
||||
_, edge_w_cfg, cent_w_cfg = _get_scoring_weights()
|
||||
|
||||
try:
|
||||
type_weight = float(payload.get("retriever_weight", 1.0))
|
||||
except (TypeError, ValueError):
|
||||
type_weight = 1.0
|
||||
|
||||
type_weight = float(payload.get("retriever_weight", 1.0))
|
||||
status_mult = _get_status_multiplier(payload)
|
||||
note_type = payload.get("type", "unknown")
|
||||
|
||||
# Breakdown für Explanation (Muss die Scoring Formel spiegeln)
|
||||
config_w_impact = type_weight - 1.0
|
||||
dynamic_b_impact = (edge_w_cfg * edge_bonus) + (cent_w_cfg * cent_bonus)
|
||||
base_val = semantic_score * status_mult
|
||||
|
||||
breakdown = ScoreBreakdown(
|
||||
semantic_contribution=(sem_w * semantic_score * type_weight * status_mult),
|
||||
edge_contribution=(edge_w_cfg * edge_bonus),
|
||||
centrality_contribution=(cent_w_cfg * cent_bonus),
|
||||
semantic_contribution=base_val,
|
||||
edge_contribution=base_val * dynamic_b_impact,
|
||||
centrality_contribution=0.0, # In dynamic_b_impact enthalten
|
||||
raw_semantic=semantic_score,
|
||||
raw_edge_bonus=edge_bonus,
|
||||
raw_centrality=cent_bonus,
|
||||
@@ -185,9 +187,8 @@ def _build_explanation(
|
||||
|
||||
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}'.", score_impact=(sem_w * semantic_score * (type_weight - 1.0))))
|
||||
reasons.append(Reason(kind="type", message=f"{msg} aufgrund des Typs '{note_type}'.", score_impact=base_val * config_w_impact))
|
||||
|
||||
# WP-22: Status Grund hinzufügen
|
||||
if status_mult != 1.0:
|
||||
msg = "Status-Bonus" if status_mult > 1.0 else "Status-Malus"
|
||||
reasons.append(Reason(kind="lifecycle", message=f"{msg} ({payload.get('status', 'unknown')}).", score_impact=0.0))
|
||||
@@ -219,13 +220,13 @@ def _build_explanation(
|
||||
reasons.append(Reason(kind="edge", message=f"{dir_txt} '{tgt_txt}' via '{top_edge.kind}'", score_impact=impact, details={"kind": top_edge.kind}))
|
||||
|
||||
if cent_bonus > 0.01:
|
||||
reasons.append(Reason(kind="centrality", message="Knoten liegt zentral im Kontext.", score_impact=breakdown.centrality_contribution))
|
||||
reasons.append(Reason(kind="centrality", message="Knoten liegt zentral im Kontext.", score_impact=cent_w_cfg * cent_bonus))
|
||||
|
||||
return Explanation(breakdown=breakdown, reasons=reasons, related_edges=edges_dto if edges_dto else None)
|
||||
|
||||
|
||||
def _extract_expand_options(req: QueryRequest) -> Tuple[int, List[str] | None]:
|
||||
"""Extrahiert depth und edge_types für die Graph-Expansion."""
|
||||
"""Extrahiert depth und edge_types für die Expansion."""
|
||||
expand = getattr(req, "expand", None)
|
||||
if not expand:
|
||||
return 0, None
|
||||
@@ -258,7 +259,7 @@ def _build_hits_from_semantic(
|
||||
explain: bool = False,
|
||||
dynamic_edge_boosts: Dict[str, float] = None
|
||||
) -> QueryResponse:
|
||||
"""Baut strukturierte QueryHits basierend auf den berechneten Scores."""
|
||||
"""Baut strukturierte QueryHits basierend auf Hybrid-Scoring."""
|
||||
t0 = time.time()
|
||||
enriched: List[Tuple[str, float, Dict[str, Any], float, float, float]] = []
|
||||
|
||||
@@ -277,27 +278,27 @@ def _build_hits_from_semantic(
|
||||
except Exception:
|
||||
cent_bonus = 0.0
|
||||
|
||||
total, edge_bonus, cent_bonus = _compute_total_score(
|
||||
total, eb, cb = _compute_total_score(
|
||||
semantic_score,
|
||||
payload,
|
||||
edge_bonus=edge_bonus,
|
||||
cent_bonus=cent_bonus,
|
||||
edge_bonus_raw=edge_bonus,
|
||||
cent_bonus_raw=cent_bonus,
|
||||
dynamic_edge_boosts=dynamic_edge_boosts
|
||||
)
|
||||
enriched.append((pid, float(semantic_score), payload, total, edge_bonus, cent_bonus))
|
||||
enriched.append((pid, float(semantic_score), payload, total, eb, cb))
|
||||
|
||||
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:
|
||||
for pid, semantic_score, payload, total, eb, cb in limited:
|
||||
explanation_obj = None
|
||||
if explain:
|
||||
explanation_obj = _build_explanation(
|
||||
semantic_score=float(semantic_score),
|
||||
payload=payload,
|
||||
edge_bonus=edge_bonus,
|
||||
cent_bonus=cent_bonus,
|
||||
edge_bonus=eb,
|
||||
cent_bonus=cb,
|
||||
subgraph=subgraph,
|
||||
node_key=payload.get("chunk_id") or payload.get("note_id")
|
||||
)
|
||||
@@ -308,8 +309,8 @@ def _build_hits_from_semantic(
|
||||
node_id=str(pid),
|
||||
note_id=payload.get("note_id", "unknown"),
|
||||
semantic_score=float(semantic_score),
|
||||
edge_bonus=edge_bonus,
|
||||
centrality_bonus=cent_bonus,
|
||||
edge_bonus=eb,
|
||||
centrality_bonus=cb,
|
||||
total_score=total,
|
||||
paths=None,
|
||||
source={
|
||||
@@ -348,7 +349,7 @@ def hybrid_retrieve(req: QueryRequest) -> QueryResponse:
|
||||
|
||||
depth, edge_types = _extract_expand_options(req)
|
||||
|
||||
# WP-22: Dynamic Boosts aus dem Request (vom Router) (Teil C)
|
||||
# WP-22: Dynamic Boosts aus dem Request (vom Router)
|
||||
boost_edges = getattr(req, "boost_edges", {})
|
||||
|
||||
subgraph: ga.Subgraph | None = None
|
||||
|
||||
+1
-1
@@ -286,7 +286,7 @@ async def chat_endpoint(
|
||||
mode="hybrid",
|
||||
top_k=request.top_k,
|
||||
explain=request.explain,
|
||||
# WP-22: Boosts weitergeben
|
||||
# WP-22: Boosts an den Retriever weitergeben
|
||||
boost_edges=edge_boosts
|
||||
)
|
||||
retrieve_result = await retriever.search(query_req)
|
||||
|
||||
@@ -2,7 +2,7 @@
|
||||
FILE: app/services/edge_registry.py
|
||||
DESCRIPTION: Single Source of Truth für Kanten-Typen. Parst '01_User_Manual/01_edge_vocabulary.md'.
|
||||
WP-22 Teil B: Registry & Validation.
|
||||
Beachtet den dynamischen Vault-Root aus ENV oder Parameter.
|
||||
FIX: Beachtet MINDNET_VAULT_ROOT aus .env korrekt.
|
||||
"""
|
||||
import re
|
||||
import os
|
||||
@@ -27,8 +27,8 @@ class EdgeRegistry:
|
||||
if self.initialized:
|
||||
return
|
||||
|
||||
# Priorität: 1. Parameter (Test) -> 2. ENV -> 3. Default
|
||||
settings = get_settings()
|
||||
# Priorität: 1. Parameter (Test) -> 2. Config (.env) -> 3. Default
|
||||
self.vault_root = vault_root or getattr(settings, "MINDNET_VAULT_ROOT", "./vault")
|
||||
self.vocab_rel_path = os.path.join("01_User_Manual", "01_edge_vocabulary.md")
|
||||
self.unknown_log_path = "data/logs/unknown_edges.jsonl"
|
||||
@@ -67,35 +67,31 @@ class EdgeRegistry:
|
||||
clean_alias = alias.replace("`", "").lower().strip()
|
||||
self.canonical_map[clean_alias] = canonical
|
||||
|
||||
logger.info(f"EdgeRegistry loaded from {full_path}: {len(self.valid_types)} canonical types.")
|
||||
logger.info(f"EdgeRegistry loaded from {full_path}: {len(self.valid_types)} types.")
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"Failed to parse Edge Vocabulary at {full_path}: {e}")
|
||||
|
||||
def resolve(self, edge_type: str) -> str:
|
||||
"""Normalisiert Kanten-Typen via Registry oder loggt Unbekannte für Review."""
|
||||
"""Normalisiert Kanten-Typen via Registry oder loggt Unbekannte."""
|
||||
if not edge_type: return "related_to"
|
||||
|
||||
# Normalisierung (Kleinschreibung, Unterstriche)
|
||||
clean_type = edge_type.lower().strip().replace(" ", "_")
|
||||
|
||||
# 1. Lookup in Map (Canonical oder Alias)
|
||||
if clean_type in self.canonical_map:
|
||||
return self.canonical_map[clean_type]
|
||||
|
||||
# 2. Unknown Handling (Loggen aber nicht verwerfen - Learning System)
|
||||
self._log_unknown(clean_type)
|
||||
return clean_type
|
||||
|
||||
def _log_unknown(self, edge_type: str):
|
||||
"""Schreibt unbekannte Typen für späteres Review in ein Log-File."""
|
||||
"""Schreibt unbekannte Typen für Review in ein Log."""
|
||||
try:
|
||||
os.makedirs(os.path.dirname(self.unknown_log_path), exist_ok=True)
|
||||
entry = {"unknown_type": edge_type, "status": "new"}
|
||||
with open(self.unknown_log_path, "a", encoding="utf-8") as f:
|
||||
f.write(json.dumps(entry) + "\n")
|
||||
except Exception:
|
||||
pass
|
||||
pass
|
||||
|
||||
# Globale Singleton Instanz
|
||||
# Singleton Instanz
|
||||
registry = EdgeRegistry()
|
||||
Reference in new issue
Block a user