perf(cache): 修复缓存失效 Bug + 集成语义缓存 + 删除 AgenticRAG 死代码

- fix: Query Cache GET/SET Key 不匹配导致命中率始终为 0%
- fix: CACHE_MIN_SCORE 阈值 0.3 对 ChromaDB cosine distance 过于严格
- feat: 在 /rag 端点集成语义缓存（命中时跳过检索+生成，92x 加速）
- refactor: 删除 AgenticRAG 死代码路径（10 个 agentic_*.py，约 1950 行）
- cleanup: 移除 engine.py 死方法、路由死函数、初始化死代码

core/engine.py

@@ -625,7 +625,7 @@ class RAGEngine:
         elif len(conditions) > 1:
             where_filter = {"$and": conditions}
 
-        query_vector = self.embedding_model.encode(query).tolist()
+        query_vector = self._encode_cached(query).tolist()
         recall_k = RERANK_CANDIDATES if (USE_RERANK or USE_HYBRID_SEARCH) else top_k
         recall_k = max(recall_k, top_k * RECALL_MULTIPLIER)
 
@@ -811,6 +811,76 @@ class RAGEngine:
             logger.warning(f"FAQ 集合查询失败: {e}")
             return get_empty_result()
 
+    def _encode_cached(self, text):
+        """
+        带缓存的 embedding 编码
+
+        优先从 Embedding Cache（LRU）读取，未命中再调用模型编码并写入缓存。
+        支持单文本和批量文本输入。
+
+        Args:
+            text: 单个文本字符串 或 文本列表
+
+        Returns:
+            numpy 数组（单文本为一维，批量为二维）
+        """
+        import numpy as _np
+
+        # 检查 embedding 缓存是否启用（缓存配置查询结果，避免每次重复导入）
+        if not hasattr(self, '_emb_cache_enabled'):
+            self._emb_cache_enabled = True  # 默认启用
+            if CACHE_AVAILABLE:
+                try:
+                    from config import EMBEDDING_CACHE_ENABLED
+                    self._emb_cache_enabled = EMBEDDING_CACHE_ENABLED
+                except ImportError:
+                    pass
+
+        if not self._emb_cache_enabled:
+            return self.embedding_model.encode(text)
+
+        try:
+            _cache = get_cache_manager()
+        except Exception:
+            return self.embedding_model.encode(text)
+
+        # 批量输入
+        if isinstance(text, list):
+            try:
+                cached_embs, missed_indices = _cache.get_embeddings_batch(text)
+                if missed_indices:
+                    missed_texts = [text[i] for i in missed_indices]
+                    # encode(list) 始终返回 2D ndarray，直接按行索引即可
+                    new_embs = self.embedding_model.encode(missed_texts)
+                    if len(missed_indices) == 1:
+                        # 单条时 encode 可能返回 1D，需统一处理
+                        if new_embs.ndim == 1:
+                            new_embs = new_embs.reshape(1, -1)
+                    for idx, mi in enumerate(missed_indices):
+                        emb_list = new_embs[idx].tolist()
+                        cached_embs[mi] = emb_list
+                        try:
+                            _cache.set_embedding(text[mi], emb_list)
+                        except Exception:
+                            pass
+                return _np.array(cached_embs)
+            except Exception:
+                # 缓存故障时优雅降级为直接编码
+                return self.embedding_model.encode(text)
+
+        # 单文本输入
+        cached = _cache.get_embedding(text)
+        if cached is not None:
+            return _np.array(cached)
+
+        embedding = self.embedding_model.encode(text)
+        try:
+            emb_list = embedding.tolist() if hasattr(embedding, 'tolist') else list(embedding)
+            _cache.set_embedding(text, emb_list)
+        except Exception:
+            pass
+        return embedding
+
     def _search_image_chunks(self, query_vector: list, top_k: int = 5, where_filter: dict = None) -> dict:
         """
         独立检索图片切片（P0：图片独立召回通道）
@@ -1387,7 +1457,7 @@ class RAGEngine:
         if not target_collections:
             return get_empty_result()
 
-        query_vector = self.embedding_model.encode(query).tolist()
+        query_vector = self._encode_cached(query).tolist()
         # 扩大召回数量，以便过滤废止切片后仍有足够结果
         recall_k = RERANK_CANDIDATES if (USE_RERANK or USE_HYBRID_SEARCH) else top_k
         recall_k = max(recall_k, top_k * RECALL_MULTIPLIER)
@@ -1739,12 +1809,12 @@ class RAGEngine:
                 # === 高精度版：基于语义向量 ===
                 from core.mmr import mmr_rerank
 
-                # 获取查询向量
-                query_emb = np.array(self.embedding_model.encode(query))
+                # 获取查询向量（使用 embedding 缓存）
+                query_emb = np.array(self._encode_cached(query))
 
-                # 批量编码所有文档
+                # 批量编码所有文档（使用 embedding 缓存）
                 docs_list = results['documents'][0]
-                all_embeddings = self.embedding_model.encode(docs_list)
+                all_embeddings = self._encode_cached(docs_list)
 
                 # 构建候选列表
                 candidates = []
@@ -1940,104 +2010,7 @@ class RAGEngine:
                 reranked[key] = results[key]
         return reranked
 
-    # ---------------- 安全与工具 ----------------
-
-    def check_restricted_documents(self, query, allowed_levels, top_k=3, role=None, department=None):
-        if not self._initialized:
-            self.initialize()
-
-        if USE_MULTI_KB and self.kb_manager and role and department:
-            from auth.gateway import get_accessible_collections
-            all_colls = [c.name for c in self.kb_manager.list_collections()]
-            accessible = set(get_accessible_collections(role, department, 'read'))
-            restricted = set(all_colls) - accessible
-
-            if not restricted:
-                return {"has_restricted": False, "restricted_levels": [], "restricted_sources": []}
-
-            query_vector = self.embedding_model.encode(query).tolist()
-            found_sources = set()
-            top_score = 0.0
-
-            for coll_name in restricted:
-                try:
-                    coll = self.kb_manager.get_collection(coll_name)
-                    if not coll: continue
-                    res = coll.query(query_embeddings=[query_vector], n_results=top_k)
-                    if res['metadatas'] and res['metadatas'][0]:
-                        for meta in res['metadatas'][0]:
-                            found_sources.add(meta.get('source', '未知'))
-                        for dist in (res.get('distances', [[]])[0] or []):
-                            if dist > top_score: top_score = dist
-                except Exception as e:
-                    logger.debug(f"权限检查遍历失败: {e}")
-
-            return {
-                "has_restricted": len(found_sources) > 0,
-                "restricted_levels": [c.replace('dept_', '') for c in restricted if True][:3],
-                "restricted_sources": list(found_sources)[:3],
-                "top_restricted_score": top_score
-            }
-
-        if not allowed_levels:
-            return {"has_restricted": False, "restricted_levels": [], "restricted_sources": [], "top_restricted_score": 0.0}
-
-        restricted_levels = {"public", "internal", "confidential", "secret"} - set(allowed_levels)
-        if not restricted_levels:
-            return {"has_restricted": False, "restricted_levels": [], "restricted_sources": [], "top_restricted_score": 0.0}
-
-        query_vector = self.embedding_model.encode(query).tolist()
-        try:
-            res = self.collection.query(
-                query_embeddings=[query_vector],
-                n_results=top_k,
-                where={"security_level": {"$in": list(restricted_levels)}}
-            )
-            docs = res.get('documents', [[]])[0]
-            if not docs:
-                return {"has_restricted": False, "restricted_levels": [], "restricted_sources": [], "top_restricted_score": 0.0}
-
-            metas = res.get('metadatas', [[]])[0]
-            dists = res.get('distances', [[]])[0]
-            found_levels, found_sources, top_score = set(), set(), 0.0
-
-            for meta, dist in zip(metas, dists):
-                found_levels.add(meta.get('security_level', 'public'))
-                found_sources.add(meta.get('source', '未知'))
-                if dist > top_score: top_score = dist
-
-            return {
-                "has_restricted": True,
-                "restricted_levels": list(found_levels),
-                "restricted_sources": list(found_sources)[:3],
-                "top_restricted_score": top_score
-            }
-        except Exception as e:
-            logger.warning(f"受限内容检查失败: {e}")
-            return {"has_restricted": False, "restricted_levels": [], "restricted_sources": [], "top_restricted_score": 0.0}
-
-    def generate_answer(self, query, context):
-        """底层生成答复能力"""
-        prompt = f"""你是一个严谨的智能助手，请根据以下参考资料回答用户的问题。
-...
-参考资料：
-{context}
-
-用户问题：{query}
-
-请回答："""
-        try:
-            from core.llm_utils import call_llm
-            result = call_llm(
-                self.llm_client,
-                prompt,
-                MODEL,
-                temperature=LLM_TEMPERATURE,
-                max_tokens=LLM_MAX_TOKENS
-            )
-            return result or f"调用大模型失败: 返回结果为空"
-        except Exception as e:
-            return f"调用大模型失败: {str(e)}"
+    # ---------------- 流式生成 ----------------
 
     def generate_answer_stream(self, query, context, history=None):
         """