fix(backend): 推荐缓存即时失效 + 相似/查重/引用补强排除 rejected + cluster_map 异步缓存

apps/api/routers/graph.py

@@ -4,7 +4,8 @@
 
 from uuid import UUID
 
-from fastapi import APIRouter, Query
+from fastapi import APIRouter, HTTPException, Query
+from fastapi.concurrency import run_in_threadpool
 
 from apps.api.deps import cache, get_paper_title, graph_service
 from packages.domain.task_tracker import global_tracker
@@ -100,30 +101,48 @@ def _fn(progress_callback=None):
 
 
 @router.get("/graph/similarity-map")
-def similarity_map(
+async def similarity_map(
     topic_id: str | None = None,
     limit: int = Query(default=200, ge=5, le=500),
 ) -> dict:
-    """论文相似度 2D 散点图（UMAP 降维）"""
-    return graph_service.similarity_map(topic_id=topic_id, limit=limit)
+    """论文相似度 2D 散点图（UMAP 降维，60s 缓存，线程池执行避免阻塞事件循环）"""
+    cache_key = f"graph_similarity_map_{topic_id}_{limit}"
+    cached = cache.get(cache_key)
+    if cached is not None:
+        return cached
+    result = await run_in_threadpool(graph_service.similarity_map, topic_id=topic_id, limit=limit)
+    cache.set(cache_key, result, ttl=60)
+    return result
 
 
 @router.get("/graph/cluster-map")
-def cluster_map(
+async def cluster_map(
     n_clusters: int = Query(default=12, ge=2, le=30),
     limit: int = Query(default=5000, ge=10, le=20000),
 ) -> dict:
-    """全库论文 embedding k-means 聚类（研究领域地图，每簇用 keywords 自动命名）"""
-    return graph_service.cluster_map(n_clusters=n_clusters, limit=limit)
+    """全库论文 embedding k-means 聚类（研究领域地图，60s 缓存，线程池执行）"""
+    cache_key = f"graph_cluster_map_{n_clusters}_{limit}"
+    cached = cache.get(cache_key)
+    if cached is not None:
+        return cached
+    result = await run_in_threadpool(graph_service.cluster_map, n_clusters=n_clusters, limit=limit)
+    cache.set(cache_key, result, ttl=60)
+    return result
 
 
 @router.get("/graph/similar-via-citation/{paper_id}")
-def similar_via_citation(
-    paper_id: str,
+async def similar_via_citation(
+    paper_id: UUID,
     top_k: int = Query(default=5, ge=1, le=20),
 ) -> dict:
     """引用同一篇论文且语义相近的论文（co-citation + 向量补强）"""
-    return graph_service.similar_via_citation(paper_id=paper_id, top_k=top_k)
+    try:
+        return await run_in_threadpool(
+            graph_service.similar_via_citation, paper_id=str(paper_id), top_k=top_k
+        )
+    except ValueError as exc:
+        # get_by_id 在论文不存在时抛 ValueError，统一转 404（此前返回 500）
+        raise HTTPException(status_code=404, detail=str(exc)) from exc
 
 
 @router.get("/graph/citation-tree/{paper_id}")

apps/api/routers/papers.py

@@ -290,6 +290,8 @@ def toggle_favorite(paper_id: UUID) -> dict:
 @router.patch("/papers/{paper_id}/reject")
 def toggle_reject(paper_id: UUID) -> dict:
     """切换论文"不感兴趣"状态（推荐系统负反馈）"""
+    from packages.ai.recommendation_service import invalidate_recommendations
+
     with session_scope() as session:
         repo = PaperRepository(session)
         try:
@@ -298,7 +300,10 @@ def toggle_reject(paper_id: UUID) -> dict:
             raise HTTPException(status_code=404, detail=str(exc)) from exc
         p.rejected = not getattr(p, "rejected", False)
         session.commit()
+        # 失效 folder_stats（左侧文件夹计数）+ 推荐缓存（被拒论文立即从推荐列表移除，
+        # 否则 recommend:{top_k} / today_summary 最长 5min 仍含该论文）
         cache.invalidate("folder_stats")
+        invalidate_recommendations()
         return {"id": str(p.id), "rejected": p.rejected}
 
 
@@ -548,7 +553,11 @@ def similar(
     paper_id: UUID,
     top_k: int = Query(default=5, ge=1, le=20),
 ) -> dict:
-    ids = rag_service.similar_papers(paper_id, top_k=top_k)
+    try:
+        ids = rag_service.similar_papers(paper_id, top_k=top_k)
+    except ValueError as exc:
+        # get_by_id 在论文不存在时抛 ValueError，统一转 404（此前返回 500）
+        raise HTTPException(status_code=404, detail=str(exc)) from exc
     items = []
     if ids:
         with session_scope() as session:
@@ -589,7 +598,11 @@ def paper_duplicates(
     from packages.ai.pipelines import PaperPipelines
 
     pipelines = PaperPipelines()
-    return pipelines.detect_duplicates(paper_id, threshold=threshold)
+    try:
+        return pipelines.detect_duplicates(paper_id, threshold=threshold)
+    except ValueError as exc:
+        # get_by_id 在论文不存在时抛 ValueError，统一转 404（此前返回 500）
+        raise HTTPException(status_code=404, detail=str(exc)) from exc
 
 
 @router.post("/papers/{paper_id}/reasoning")

packages/ai/graph/similarity.py

@@ -185,8 +185,6 @@ def similar_via_citation(self, paper_id: str, top_k: int = 5) -> dict:
         with session_scope() as session:
             repo = PaperRepository(session)
             seed = repo.get_by_id(paper_id)
-            if not seed:
-                return {"paper_id": str(paper_id), "items": [], "note": "论文不存在"}
             if not seed.embedding:
                 return {
                     "paper_id": str(paper_id),
@@ -195,13 +193,15 @@ def similar_via_citation(self, paper_id: str, top_k: int = 5) -> dict:
                 }
             seed_vec = list(seed.embedding)
 
-            # 1. 取种子论文引用 / 被引的论文（双向）
+            # 1. 取种子论文引用 / 被引的论文（双向，限 200 防高被引 hub 过载）
             cited_rows = (
                 session.execute(
-                    _sa_select(Citation).where(
+                    _sa_select(Citation)
+                    .where(
                         (Citation.source_paper_id == str(paper_id))
                         | (Citation.target_paper_id == str(paper_id))
                     )
+                    .limit(200)
                 )
                 .scalars()
                 .all()
@@ -215,13 +215,15 @@ def similar_via_citation(self, paper_id: str, top_k: int = 5) -> dict:
             if not cited_ids:
                 return {"paper_id": str(paper_id), "items": [], "note": "无引用关系"}
 
-            # 2. 找也引用了 cited 中任一篇的论文（co-citation）
+            # 2. 找也引用了 cited 中任一篇的论文（co-citation，限 2000 防无界扩张）
             co_cite_rows = (
                 session.execute(
-                    _sa_select(Citation).where(
+                    _sa_select(Citation)
+                    .where(
                         Citation.source_paper_id.in_(list(cited_ids))
                         | Citation.target_paper_id.in_(list(cited_ids))
                     )
+                    .limit(2000)
                 )
                 .scalars()
                 .all()
@@ -236,13 +238,16 @@ def similar_via_citation(self, paper_id: str, top_k: int = 5) -> dict:
             if not co_cite_ids:
                 return {"paper_id": str(paper_id), "items": [], "note": "无 co-citation 候选"}
 
-            # 3. 在 co-citation 集合里按 embedding cosine 排序
+            # 3. 在 co-citation 集合里按 embedding cosine 排序（排除被屏蔽论文）
             candidates = (
                 session.execute(
-                    _sa_select(Paper).where(
+                    _sa_select(Paper)
+                    .where(
                         Paper.id.in_(list(co_cite_ids)),
                         Paper.embedding.is_not(None),
+                        Paper.rejected.is_(False),
                     )
+                    .limit(500)
                 )
                 .scalars()
                 .all()

packages/ai/recommendation_service.py

@@ -41,6 +41,17 @@ def _set_cache(key: str, value: object):
         _ttl_cache[key] = (time.monotonic(), value)
 
 
+def invalidate_recommendations() -> None:
+    """清空推荐与今日简报缓存（toggle_reject 后调用，使负反馈立即生效）。
+
+    覆盖 `recommend:{top_k}` 与 `today_summary`（简报内嵌 recommend(top_k=5)）。
+    与 deps.cache 是两套独立 dict，不能复用 deps.cache.invalidate。
+    """
+    with _ttl_lock:
+        for k in [k for k in _ttl_cache if k.startswith("recommend:") or k == "today_summary"]:
+            del _ttl_cache[k]
+
+
 def _mean_vector(vectors: list[list[float]]) -> list[float]:
     """计算向量集合的质心（自动过滤维度不一致的向量）"""
     if not vectors:

packages/storage/repositories/paper.py

@@ -392,6 +392,7 @@ def similar_by_embedding(
                 select(Paper)
                 .where(Paper.id != str(exclude))
                 .where(Paper.embedding.is_not(None))
+                .where(Paper.rejected.is_(False))
                 .order_by(Paper.embedding.cosine_distance(vector))
                 .limit(limit)
             )
@@ -401,6 +402,7 @@ def similar_by_embedding(
             select(Paper)
             .where(Paper.id != str(exclude))
             .where(Paper.embedding.is_not(None))
+            .where(Paper.rejected.is_(False))
             .order_by(Paper.created_at.desc())
             .limit(max_candidates)
         )
@@ -438,6 +440,7 @@ def semantic_candidates(
             q = (
                 select(Paper)
                 .where(Paper.embedding.is_not(None))
+                .where(Paper.rejected.is_(False))
                 .order_by(Paper.embedding.cosine_distance(query_vector))
                 .limit(limit)
             )
@@ -446,6 +449,7 @@ def semantic_candidates(
         q = (
             select(Paper)
             .where(Paper.embedding.is_not(None))
+            .where(Paper.rejected.is_(False))
             .order_by(Paper.created_at.desc())
             .limit(max_candidates)
         )