app/core/retriever.py aktualisiert

app/core/retriever.py

@@ -2,21 +2,24 @@
 app/core/retriever.py — Semantischer/Edge-Aware/Hybrid Retriever (WP-04)
 
 Zweck:
-    Kandidatenfindung via Vektorsuche in *_chunks, optionale Edge-Expansion und
-    kombiniertes Ranking zur Rückgabe von Top-K Treffern.
+    Kandidatenfindung via Vektorsuche in *_chunks, optionale Edge-Expansion
+    und kombiniertes Ranking zur Rückgabe von Top-K Treffern.
+    Erweiterung (0.2.0): Text→Embedding, falls kein query_vector übergeben wurde.
+
 Kompatibilität:
     Python 3.12+, qdrant-client 1.x
 Version:
-    0.1.0  (Erstanlage)
+    0.2.0  (Text→Embedding ergänzt; bestehendes Verhalten unverändert)
 Stand:
     2025-10-07
 Bezug:
     - app/core/graph_adapter.py (expand)
     - app/core/ranking.py (combine_scores)
     - app/core/qdrant_points.py (search_chunks_by_vector)
-Nutzung:
-    from app.core.retriever import hybrid_retrieve
+    - app/services/embeddings_client.py (embed_text)
+    - app/models/dto.py (QueryRequest/Response)
 Änderungsverlauf:
+    0.2.0 (2025-10-07) – Text→Embedding (embed_text_if_needed).
     0.1.0 (2025-10-07) – Erstanlage.
 """
 
@@ -30,18 +33,21 @@ from app.core.ranking import combine_scores
 from app.core.graph_adapter import expand
 from app.core import qdrant_points as qp
 from app.config import get_settings
+from app.services.embeddings_client import embed_text
 
 
-def _require_query_vector(req: QueryRequest) -> List[float]:
+def _vector_from_request(req: QueryRequest) -> List[float]:
     """
-    Für den Schnelltest ohne eingebundene Embeddings muss query_vector gesetzt sein.
-    Später kann hier der Embed-Aufruf (Text → 384d) angebunden werden.
+    Query-Vektor bestimmen:
+      - Falls query_vector gesetzt: unverändert verwenden (Back-compat, Tests).
+      - Sonst, falls query gesetzt: serverseitig einbetten.
+      - Andernfalls: Fehler.
     """
-    if not req.query_vector:
-        raise ValueError(
-            "query_vector fehlt. Für den Quick-Test ohne Embeddings bitte einen 384d-Vektor übergeben."
-        )
-    return req.query_vector
+    if req.query_vector:
+        return req.query_vector
+    if req.query:
+        return embed_text(req.query)
+    raise ValueError("query_vector fehlt. Alternativ 'query' (Text) übergeben, wird serverseitig eingebettet.")
 
 
 def semantic_retrieve(req: QueryRequest) -> QueryResponse:
@@ -50,9 +56,8 @@ def semantic_retrieve(req: QueryRequest) -> QueryResponse:
     s = get_settings()
     client = QdrantClient(url=s.QDRANT_URL, api_key=s.QDRANT_API_KEY)
 
-    q_vec = _require_query_vector(req)
+    q_vec = _vector_from_request(req)
     raw_hits = qp.search_chunks_by_vector(client, s.COLLECTION_PREFIX, q_vec, top=req.top_k, filters=req.filters)
-    id2payload = {pid: payload for (pid, score, payload) in raw_hits}
 
     results: List[QueryHit] = []
     for pid, s_score, payload in raw_hits:
@@ -62,11 +67,10 @@ def semantic_retrieve(req: QueryRequest) -> QueryResponse:
             semantic_score=float(s_score),
             edge_bonus=0.0,
             centrality_bonus=0.0,
-            total_score=float(s_score),  # hier un-normalisiert; ok für schnelle Prüfung
+            total_score=float(s_score),  # un-normalisiert: ok für quick semantic mode
             paths=None,
             source={"path": payload.get("path"), "section": payload.get("section_title")}
         ))
-
     dt = int((time.time() - t0) * 1000)
     return QueryResponse(results=results, used_mode="semantic", latency_ms=dt)
 
@@ -77,11 +81,11 @@ def hybrid_retrieve(req: QueryRequest) -> QueryResponse:
     s = get_settings()
     client = QdrantClient(url=s.QDRANT_URL, api_key=s.QDRANT_API_KEY)
 
-    q_vec = _require_query_vector(req)
+    q_vec = _vector_from_request(req)
 
     # 1) Semantische Seeds (top_k * 3 für breitere Basis)
     raw_hits = qp.search_chunks_by_vector(client, s.COLLECTION_PREFIX, q_vec, top=req.top_k * 3, filters=req.filters)
-    id2payload = {pid: payload for (pid, score, payload) in raw_hits}
+    id2payload = {pid: payload for (pid, _, payload) in raw_hits}
     seeds = [pid for (pid, _, _) in raw_hits]
 
     # 2) Edge-Expansion
@@ -93,10 +97,10 @@ def hybrid_retrieve(req: QueryRequest) -> QueryResponse:
     centrality_map = {pid: sg.centrality_bonus(pid) for pid in seeds}
 
     # 3) Combined Ranking
-    scored = combine_scores(raw_hits, edge_bonus_map, centrality_map,
-                            w_sem=s.RETRIEVER_W_SEM,
-                            w_edge=s.RETRIEVER_W_EDGE,
-                            w_cent=s.RETRIEVER_W_CENT)
+    scored = combine_scores(
+        raw_hits, edge_bonus_map, centrality_map,
+        w_sem=s.RETRIEVER_W_SEM, w_edge=s.RETRIEVER_W_EDGE, w_cent=s.RETRIEVER_W_CENT
+    )
 
     # 4) Antwortobjekte (Chunk-Ebene)
     results: List[QueryHit] = []