feat(multimodal): 实现本地模型管线与请求用量配置

backend/app/retrieval/embedding.py

@@ -1,7 +1,6 @@
 """Embedding 统一接口与轻量实现。
 
-真实默认是本地 BGE-M3 类模型，但第一阶段先跑通链路，这里用确定性的特征哈希向量代替。
-后续接入真实模型时实现同样的 EmbeddingProvider 接口替换即可，上层检索逻辑不变。
+生产环境使用 local_models 的真实模型。特征哈希实现仅供测试显式注入。
 """
 
 from __future__ import annotations

backend/app/retrieval/engine.py

@@ -20,6 +20,7 @@ from app.contracts import (
 )
 from app.repository import BlockHit
 from app.retrieval.embedding import EmbeddingProvider, HashEmbeddingProvider
+from app.local_models.runtime import LocalEmbedding
 from app.retrieval.hybrid import normalize_scores, rrf_fuse
 from app.retrieval.reranker import LexicalReranker, RankedCandidate, RerankerProvider
 from app.retrieval import routed_vectors
@@ -88,8 +89,13 @@ class RetrievalEngine:
                 and self.embedding is self._routed_defaults[0]
                 and self.vector_store is self._routed_defaults[1]
             ):
-                vec_hits = await routed_vectors.search_remote(request.query, top_k=recall)
+                vec_hits = await routed_vectors.search_remote(request.query, top_k=recall, accept_local=isinstance(self.embedding, LocalEmbedding))
             if vec_hits is None:
+                if isinstance(self.embedding, LocalEmbedding):
+                    if request.mode == SearchMode.hybrid:
+                        return self._search_fts(request)
+                    from app.errors import ApiError
+                    raise ApiError(409, "SEMANTIC_INDEX_UNAVAILABLE", "语义索引未就绪。请配置 Embedding 或下载本地模型后重建索引。")
                 query_vec = await self.embedding.embed_query(request.query)
                 vec_hits = await self.vector_store.search(query_vec, top_k=recall)
                 record_embedding(source="local", model_id=self.embedding.model_id,
@@ -287,5 +293,5 @@ def _utc(dt: datetime) -> datetime:
 
 # 默认引擎实例：轻量实现跑通链路，后续可替换真实模型实现
 engine = RetrievalEngine(
-    HashEmbeddingProvider(), LexicalReranker(), SqliteVecStore(), route_embeddings=True,
+    LocalEmbedding(), LexicalReranker(), SqliteVecStore(), route_embeddings=True,
 )

backend/app/retrieval/routed_vectors.py

@@ -42,6 +42,7 @@ class RemoteEmbeddings:
     space_id: str
     dimensions: int
     vectors: list[list[float]]
+    source: str = "api"
 
 
 def get_model_routing() -> EmbeddingRuntime | None:
@@ -67,7 +68,7 @@ def _unit_vector(vector: list[float], dimensions: int) -> list[float]:
     return [value / norm for value in scaled]
 
 
-async def embed_remote(texts: list[str]) -> RemoteEmbeddings | None:
+async def embed_remote(texts: list[str], *, accept_local=False) -> RemoteEmbeddings | None:
     """Return validated API vectors, or None to use the caller's local baseline.
 
     Do not use the runtime's local result: the caller may have injected its own
@@ -80,7 +81,7 @@ async def embed_remote(texts: list[str]) -> RemoteEmbeddings | None:
         if runtime is None:
             return None
         result = await runtime.embed(texts)
-        if result.source != "api":
+        if result.source != "api" and not accept_local:
             record_embedding(fallback_reason=result.fallback_reason)
             return None
         if not isinstance(result.model_id, str) or not result.model_id or result.model_id == "hash-v1":
@@ -93,6 +94,7 @@ async def embed_remote(texts: list[str]) -> RemoteEmbeddings | None:
             space_id=result.model_id,
             dimensions=result.dimensions,
             vectors=[_unit_vector(vector, result.dimensions) for vector in result.vectors],
+            source=result.source,
         )
     except Exception as exc:
         # Avoid logging provider exceptions containing credentials or note text.
@@ -152,13 +154,13 @@ def store_remote(
         logger.warning("Remote vector storage unavailable (%s); local index retained", type(exc).__name__)
 
 
-async def search_remote(query: str, *, top_k: int) -> list[VectorHit] | None:
+async def search_remote(query: str, *, top_k: int, accept_local=False) -> list[VectorHit] | None:
     """None means fallback, including any missing/invalid current-block vector.
 
     Read coverage and vectors together so concurrent note updates cannot produce
     an apparently complete subset. Never fill missing remote hits with local hits.
     """
-    batch = await embed_remote([query])
+    batch = await embed_remote([query], accept_local=accept_local)
     if batch is None:
         return None
     record_embedding(attempted_space={"model_id": batch.space_id, "dimensions": batch.dimensions})
@@ -190,7 +192,7 @@ async def search_remote(query: str, *, top_k: int) -> list[VectorHit] | None:
                         yield VectorHit(id=row["block_id"], score=max(0.0, min(1.0, score)))
 
                 result = heapq.nlargest(top_k, hits(), key=lambda hit: hit.score)
-                record_embedding(source="api", model_id=batch.space_id,
+                record_embedding(source=batch.source, model_id=batch.space_id,
                                  dimensions=batch.dimensions, fallback_reason=None)
                 return result
         finally: