음성 재생과 운영 배포 정리

apps/api/app/services/rag.py

@@ -252,6 +252,7 @@ def _vector_literal(vec: Sequence[float]) -> str:
 #   $7 = w_sparse(real)
 #   $8 = pre_k   (int — dense/sparse 각 후보 수, 보통 50)
 #   $9 = k       (int — 융합 후 반환 수)
+#   $10 = source_ids(text[] — 빈 배열이면 전체 허용)
 #
 # 정보비대칭 강제: 두 CTE 모두 `$4 = ANY(visible_to) AND sensitivity <= $5` 사전필터.
 # kinds 빈 배열 처리: cardinality($3)=0 이면 kb_kind 조건을 통과(전체).
@@ -267,6 +268,7 @@ dense AS (
   FROM kb.chunk c, params p
   WHERE c.embedding IS NOT NULL
     AND (cardinality($3::text[]) = 0 OR c.kb_kind = ANY($3::text[]))
+    AND (cardinality($10::text[]) = 0 OR c.source_id = ANY($10::text[]))
     AND $4 = ANY(c.visible_to)
     AND c.sensitivity <= $5
   ORDER BY c.embedding <=> p.q_dense
@@ -279,6 +281,7 @@ sparse AS (
   WHERE p.q_ts IS NOT NULL
     AND to_tsvector('simple', c.chunk_text) @@ p.q_ts
     AND (cardinality($3::text[]) = 0 OR c.kb_kind = ANY($3::text[]))
+    AND (cardinality($10::text[]) = 0 OR c.source_id = ANY($10::text[]))
     AND $4 = ANY(c.visible_to)
     AND c.sensitivity <= $5
   ORDER BY s_sparse DESC
@@ -473,6 +476,7 @@ async def search_kb(
         fs = filters.get("sensitivity_max")
         if isinstance(fs, int):
             sens_max = min(sens_max, fs)  # 더 엄격하게만
+    source_filter = [str(item) for item in (filters.get("source_id", []) if filters else [])]
 
     # (2) 질의 임베딩(dense+sparse). 모델 미가용 → NotConfigured 전파.
     eq = await asyncio.to_thread(embed_query, query)  # CPU 인코딩 → 스레드풀(이벤트루프 비차단)
@@ -491,17 +495,13 @@ async def search_kb(
             policy.w_sparse,             # $7
             pre_k,                       # $8 pre_k
             max(k * 4, k),               # $9 융합 후 1차 컷(리랭킹 입력 여유분)
+            source_filter,               # $10 source_id 좁힘
         )
     except Exception as e:  # UndefinedFunction(vector 미설치) / UndefinedColumn 등
         raise NotConfigured(f"KB hybrid query failed (DB/pgvector not ready): {e}") from e
 
-    # source_id 추가 좁힘(SQL 후처리 — 화이트리스트 보존, 코드 단순화)
-    src_filter = set(filters.get("source_id", [])) if filters else set()
-
     chunks: list[RetrievedChunk] = []
     for r in rows:
-        if src_filter and r["source_id"] not in src_filter:
-            continue
         # asyncpg는 jsonb를 str(JSON text)로 반환 → 파싱. 코덱 등록 시 dict 그대로도 수용.
         _meta_raw = r["meta"]
         meta = json.loads(_meta_raw) if isinstance(_meta_raw, str) else dict(_meta_raw or {})
@@ -635,6 +635,7 @@ async def retrieve_eval_grounding(
     query: str,
     k: int = 5,
     kinds: Optional[Sequence[str]] = None,
+    source_ids: Optional[Sequence[str]] = None,
     rerank: bool = True,
 ) -> RetrievalResult:
     """평가 AI 채점 근거 회수 — DSM/이론/taxonomy 정답라벨 + 논평.
@@ -644,13 +645,17 @@ async def retrieve_eval_grounding(
 
     kinds: 평가 차원에 따라 좁히기(예: 기법 채점 → ['technique','supervisor_pattern']).
     """
-    filters = {"kb_kind": list(kinds)} if kinds else None
+    filters: dict[str, Any] = {}
+    if kinds:
+        filters["kb_kind"] = list(kinds)
+    if source_ids:
+        filters["source_id"] = list(source_ids)
     return await search_kb(
         conn,
         query=query,
         role=AIRole.EVALUATOR,
         k=k,
-        filters=filters,
+        filters=filters or None,
         rerank=rerank,
     )