perf: skip CPU cross-encoder reranking by default

app/agents/nodes.py

@@ -213,15 +213,24 @@ async def retrieve_node(state: AgentState) -> dict[str, Any]:
 
 # ========== Node 4: rerank ==========
 async def rerank_node(state: AgentState) -> dict[str, Any]:
-    """BGE-reranker 精排 top-N + 产出引用."""
+    """可选 CrossEncoder 精排；CPU 默认沿用混合检索排序并产出引用。"""
     started = time.time()
     hits = state.get("retrieved") or []
     query = state.get("query_rewritten") or _last_user_query(state)
     if not hits:
         return {"reranked": [], "citations": [], "relevance_score": 0.0, "relevance_verdict": "irrelevant"}
 
-    reranker = get_reranker_service()
-    reranked = await reranker.rerank(query, hits, top_n=settings.rerank_top_n)
+    if settings.enable_reranker:
+        reranker = get_reranker_service()
+        reranked = await reranker.rerank(query, hits, top_n=settings.rerank_top_n)
+    else:
+        # BGE-M3 dense + sparse 已在 hybrid_query 中完成 RRF 排序。免费 CPU Space
+        # 上再跑 CrossEncoder 实测会额外阻塞约 73 秒；直接取 top-N，并用保留下来的
+        # dense cosine 作为可解释的相关性分数。GPU 部署可通过环境变量恢复精排。
+        reranked = list(hits[: settings.rerank_top_n])
+        for rank, hit in enumerate(reranked):
+            hit.original_rank = rank
+            hit.rerank_score = max(0.0, min(1.0, float(hit.dense_score)))
 
     # 构造引用 (前 5 个, 按 rerank 分数)
     citations: list[dict[str, Any]] = []
@@ -238,7 +247,12 @@ async def rerank_node(state: AgentState) -> dict[str, Any]:
         })
 
     top_score = reranked[0].rerank_score if reranked else 0.0
-    if top_score >= settings.crag_relevance_threshold:
+    relevance_threshold = (
+        settings.crag_relevance_threshold
+        if settings.enable_reranker
+        else settings.hybrid_relevance_threshold
+    )
+    if top_score >= relevance_threshold:
         verdict = "relevant"
     elif top_score < 0.3:
         verdict = "irrelevant"

app/config.py

@@ -61,12 +61,15 @@ class Settings(BaseSettings):
     reranker_model: str = "BAAI/bge-reranker-v2-m3"
     embedding_device: Literal["cpu", "cuda", "mps"] = "cpu"
     use_fp16: bool = True
+    # HF 免费 CPU 上对 20 个候选做 CrossEncoder 重排实测约 73 秒。
+    # 默认使用 BGE-M3 dense+sparse 混合排序；GPU Space 可设 ENABLE_RERANKER=true。
+    enable_reranker: bool = False
 
     # ========== 向量库 / ChromaDB ==========
     chroma_persist_dir: str = "./data/chroma"
     chroma_collection: str = "docs"
     # ⚡ A 改良版: 默认关闭 ColBERT. 三路融合 (dense+sparse+colbert) 每次查询
-    # 多扫 30 个 .npy 文件 + matmul, 0.3-1.5s 开销. 召回率轻微降, reranker 兜底.
+    # 多扫 30 个 .npy 文件 + matmul, 0.3-1.5s 开销. 默认保留 dense+sparse 双路融合.
     # 想恢复三路融合, 设环境变量 ENABLE_COLBERT=true.
     enable_colbert: bool = False
 
@@ -97,6 +100,7 @@ class Settings(BaseSettings):
     rerank_top_n: int = 5
     crag_max_iterations: int = 2
     crag_relevance_threshold: float = 0.7
+    hybrid_relevance_threshold: float = 0.45
 
     # ========== LLM 缓存 ==========
     llm_cache_enabled: bool = True

app/main.py

@@ -58,7 +58,7 @@ async def lifespan(app: FastAPI):
     except Exception as e:  # noqa: BLE001
         logger.warning("ChromaDB init failed (will retry on first request): %s", e)
 
-    # BGE-M3 & Reranker 预热.
+    # BGE-M3 & 可选 Reranker 预热.
     # 必须在主线程 *同步* 执行, 不能 run_in_executor 异步跑 — 因为 FlagEmbedding
     # 在 torch 2.2 + transformers 4.57 组合下, 子线程首次 .to(device) 会撞 meta tensor
     # 错 "Cannot copy out of meta tensor; no data!"; 必须等它 meta→cpu 转移完再 await
@@ -66,7 +66,8 @@ async def lifespan(app: FastAPI):
     if settings.embedding_model and not os.environ.get("TESTING"):
         try:
             warm_embedder()
-            warm_reranker()
+            if settings.enable_reranker:
+                warm_reranker()
         except Exception as e:  # noqa: BLE001
             logger.warning("Embedder/reranker warm-up failed: %s", e)
 

app/services/vector_store.py

@@ -371,6 +371,9 @@ def hybrid_query(
 
     # RRF 融合
     fused = rrf_fuse(dense, sparse, colbert_ranked, k=60)[:k]
+    dense_scores = {cid: score for cid, score, _payload in dense}
+    sparse_scores_by_id = {cid: score for cid, score, _payload in sparse}
+    colbert_scores = {cid: score for cid, score, _payload in colbert_ranked}
 
     # 构造 RetrievalHit
     hits: list[RetrievalHit] = []
@@ -385,6 +388,9 @@ def hybrid_query(
             heading=meta.get("heading"),
             context_prefix=meta.get("context_prefix"),
             meta=meta,
+            dense_score=float(dense_scores.get(cid, 0.0)),
+            sparse_score=float(sparse_scores_by_id.get(cid, 0.0)),
+            colbert_score=float(colbert_scores.get(cid, 0.0)),
         ))
 
     logger.debug("hybrid_query returned %d hits in %dms", len(hits), int((time.time() - started) * 1000))