This commit is contained in:
Lars 2025-12-18 13:21:53 +01:00
parent 3eac646cb6
commit 48729e6f5d
3 changed files with 46 additions and 44 deletions

View File

@ -162,7 +162,7 @@ class IngestionService:
# --- WP-22: Content Lifecycle Gate ---
status = fm.get("status", "draft").lower().strip()
# Hard Skip für System-Dateien (Teil A)
# Hard Skip für System-Dateien
if status in ["system", "template", "archive", "hidden"]:
logger.info(f"Skipping file {file_path} (Status: {status})")
return {**result, "status": "skipped", "reason": f"lifecycle_status_{status}"}
@ -265,7 +265,7 @@ class IngestionService:
except TypeError:
raw_edges = build_edges_for_note(note_id, chunk_pls)
# --- WP-22: Edge Registry Validation (Teil B) ---
# --- WP-22: Edge Registry Validation ---
edges = []
if raw_edges:
for edge in raw_edges:

View File

@ -98,7 +98,7 @@ def _semantic_hits(
results.append((str(pid), float(score), dict(payload or {})))
return results
# --- WP-22 Helper: Lifecycle Multipliers (Teil A) ---
# --- WP-22 Helper: Lifecycle Multipliers ---
def _get_status_multiplier(payload: Dict[str, Any]) -> float:
"""
WP-22: Drafts werden bestraft, Stable Notes belohnt.
@ -106,11 +106,10 @@ def _get_status_multiplier(payload: Dict[str, Any]) -> float:
status = str(payload.get("status", "draft")).lower()
if status == "stable": return 1.2
if status == "active": return 1.0
if status == "draft": return 0.5 # Malus für Entwürfe
if status == "draft": return 0.8 # Malus für Entwürfe
# Fallback für andere oder leere Status
return 1.0
# --- WP-22: Dynamic Scoring Formula (Teil C) ---
def _compute_total_score(
semantic_score: float,
payload: Dict[str, Any],
@ -119,8 +118,8 @@ def _compute_total_score(
dynamic_edge_boosts: Dict[str, float] = None
) -> Tuple[float, float, float]:
"""
Berechnet total_score nach WP-22 Formel.
Score = (Sem * Type * Status) + (Weighted_Edge + Cent)
Berechnet total_score.
WP-22 Update: Integration von Status-Bonus und Dynamic Edge Boosts.
"""
raw_weight = payload.get("retriever_weight", 1.0)
try:
@ -133,13 +132,13 @@ def _compute_total_score(
sem_w, edge_w, cent_w = _get_scoring_weights()
status_mult = _get_status_multiplier(payload)
# Dynamic Edge Boosting (Teil C)
# Wenn dynamische Boosts aktiv sind (durch den Router), verstärken wir den Graph-Bonus global.
# Der konkrete kanten-spezifische Boost passiert bereits im Subgraph (hybrid_retrieve).
# Dynamic Edge Boosting
# Wenn dynamische Boosts aktiv sind, erhöhen wir den Einfluss des Graphen
# Dies ist eine Vereinfachung, da der echte Boost im Subgraph passiert sein sollte.
final_edge_score = edge_w * edge_bonus
if dynamic_edge_boosts and edge_bonus > 0:
# Globaler Boost-Faktor falls Intention (z.B. WHY) vorliegt
final_edge_score *= 1.5
# Globaler Boost für Graph-Signale bei spezifischen Intents
final_edge_score *= 1.2
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)
@ -155,8 +154,9 @@ def _build_explanation(
subgraph: Optional[ga.Subgraph],
node_key: Optional[str]
) -> Explanation:
"""Erstellt ein Explanation-Objekt (WP-04b)."""
"""Erstellt ein Explanation-Objekt."""
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:
@ -167,7 +167,6 @@ def _build_explanation(
status_mult = _get_status_multiplier(payload)
note_type = payload.get("type", "unknown")
# Breakdown Berechnung (muss mit _compute_total_score korrelieren)
breakdown = ScoreBreakdown(
semantic_contribution=(sem_w * semantic_score * type_weight * status_mult),
edge_contribution=(edge_w_cfg * edge_bonus),
@ -181,7 +180,6 @@ def _build_explanation(
reasons: List[Reason] = []
edges_dto: List[EdgeDTO] = []
# Reason Generation Logik (WP-04b)
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:
@ -191,13 +189,11 @@ def _build_explanation(
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))))
# NEU: WP-22 Status Reason
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')}).", score_impact=0.0))
if subgraph and node_key and edge_bonus > 0:
# Extrahiere Top-Kanten für die Erklärung
if hasattr(subgraph, "get_outgoing_edges"):
outgoing = subgraph.get_outgoing_edges(node_key)
for edge in outgoing:
@ -230,7 +226,7 @@ def _build_explanation(
def _extract_expand_options(req: QueryRequest) -> Tuple[int, List[str] | None]:
"""Extrahiert depth und edge_types für Graph-Expansion."""
"""Extrahiert depth und edge_types."""
expand = getattr(req, "expand", None)
if not expand:
return 0, None
@ -263,7 +259,7 @@ def _build_hits_from_semantic(
explain: bool = False,
dynamic_edge_boosts: Dict[str, float] = None
) -> QueryResponse:
"""Baut strukturierte QueryHits basierend auf Scoring (WP-22 & WP-04b)."""
"""Baut strukturierte QueryHits."""
t0 = time.time()
enriched: List[Tuple[str, float, Dict[str, Any], float, float, float]] = []
@ -282,28 +278,27 @@ def _build_hits_from_semantic(
except Exception:
cent_bonus = 0.0
total, eb, cb = _compute_total_score(
total, edge_bonus, cent_bonus = _compute_total_score(
semantic_score,
payload,
edge_bonus=edge_bonus,
cent_bonus=cent_bonus,
dynamic_edge_boosts=dynamic_edge_boosts
)
enriched.append((pid, float(semantic_score), payload, total, eb, cb))
enriched.append((pid, float(semantic_score), payload, total, edge_bonus, cent_bonus))
# Sort & Limit
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, eb, cb in limited:
for pid, semantic_score, payload, total, edge_bonus, cent_bonus in limited:
explanation_obj = None
if explain:
explanation_obj = _build_explanation(
semantic_score=float(semantic_score),
payload=payload,
edge_bonus=eb,
cent_bonus=cb,
edge_bonus=edge_bonus,
cent_bonus=cent_bonus,
subgraph=subgraph,
node_key=payload.get("chunk_id") or payload.get("note_id")
)
@ -312,10 +307,10 @@ def _build_hits_from_semantic(
results.append(QueryHit(
node_id=str(pid),
note_id=payload.get("note_id", "unknown"),
note_id=payload.get("note_id"),
semantic_score=float(semantic_score),
edge_bonus=eb,
centrality_bonus=cb,
edge_bonus=edge_bonus,
centrality_bonus=cent_bonus,
total_score=total,
paths=None,
source={
@ -332,7 +327,7 @@ def _build_hits_from_semantic(
def semantic_retrieve(req: QueryRequest) -> QueryResponse:
"""Reiner semantischer Retriever (WP-02)."""
"""Reiner semantischer Retriever."""
client, prefix = _get_client_and_prefix()
vector = _get_query_vector(req)
top_k = req.top_k or get_settings().RETRIEVER_TOP_K
@ -342,44 +337,44 @@ def semantic_retrieve(req: QueryRequest) -> QueryResponse:
def hybrid_retrieve(req: QueryRequest) -> QueryResponse:
"""Hybrid-Retriever: semantische Suche + optionale Edge-Expansion (WP-04a)."""
"""Hybrid-Retriever: semantische Suche + optionale Edge-Expansion."""
client, prefix = _get_client_and_prefix()
if req.query_vector:
vector = list(req.query_vector)
else:
vector = _get_query_vector(req)
# 1. Semantische Suche
vector = list(req.query_vector) if req.query_vector else _get_query_vector(req)
top_k = req.top_k or get_settings().RETRIEVER_TOP_K
hits = _semantic_hits(client, prefix, vector, top_k=top_k, filters=req.filters)
# 2. Graph Expansion & Custom Boosting (WP-22 Teil C)
depth, edge_types = _extract_expand_options(req)
# WP-22: Dynamic Boosts aus dem Request (vom Router)
boost_edges = getattr(req, "boost_edges", {})
subgraph: ga.Subgraph | None = None
if depth and depth > 0:
seed_ids: List[str] = []
for _pid, _score, payload in hits:
key = payload.get("note_id")
key = payload.get("chunk_id") or payload.get("note_id")
if key and key not in seed_ids:
seed_ids.append(key)
if seed_ids:
try:
# Subgraph laden
# Hier könnten wir boost_edges auch an expand übergeben, wenn ga.expand es unterstützt
subgraph = ga.expand(client, prefix, seed_ids, depth=depth, edge_types=edge_types)
# --- WP-22: Kanten-Boosts im RAM-Graphen anwenden ---
# Dies manipuliert die Gewichte im Graphen, bevor der 'edge_bonus' berechnet wird.
# Manuelles Boosten der Kantengewichte im Graphen falls aktiv
if boost_edges and subgraph and hasattr(subgraph, "graph"):
for u, v, data in subgraph.graph.edges(data=True):
k = data.get("kind")
if k in boost_edges:
# Gewicht multiplizieren (z.B. caused_by * 3.0)
# Gewicht erhöhen für diesen Query-Kontext
data["weight"] = data.get("weight", 1.0) * boost_edges[k]
except Exception:
subgraph = None
# 3. Scoring & Re-Ranking
return _build_hits_from_semantic(
hits,
top_k=top_k,
@ -391,6 +386,11 @@ def hybrid_retrieve(req: QueryRequest) -> QueryResponse:
class Retriever:
"""Wrapper-Klasse für Suchoperationen."""
"""
Wrapper-Klasse für WP-05 (Chat).
"""
def __init__(self):
pass
async def search(self, request: QueryRequest) -> QueryResponse:
return hybrid_retrieve(request)

View File

@ -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: Dynamische Pfad-Auflösung basierend auf MINDNET_VAULT_ROOT.
"""
import re
import os
@ -25,7 +25,7 @@ class EdgeRegistry:
if self.initialized:
return
# Priorität: 1. Parameter -> 2. ENV -> 3. Default
# Priorität: 1. Parameter (Test) -> 2. ENV (dotenv) -> 3. Default
self.vault_root = vault_root or os.getenv("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"
@ -38,9 +38,11 @@ class EdgeRegistry:
def _load_vocabulary(self):
"""Parst die Markdown-Tabelle im Vault."""
# Absoluten Pfad auflösen, um Verwechslungen im venv zu vermeiden
full_path = os.path.abspath(os.path.join(self.vault_root, self.vocab_rel_path))
if not os.path.exists(full_path):
# Debug-Info: Zeige wo genau gesucht wurde
logger.warning(f"Edge Vocabulary NOT found at: {full_path}. Registry is empty.")
return