fix: search_hybrid 改用改写后的 retrieval_query + 清理 engine.py 死代码

关键修复:
- chat_routes.py: search_hybrid(message,...) → search_hybrid(retrieval_query,...)
  修复无状态模式下追问场景("那工作流程呢？")用原始模糊文本做向量检索的问题，
  现在使用意图分析改写后的完整查询("货源投放的具体工作流程是什么？")

死代码清理 (Claude Code review 反馈):
- engine.py: 移除 4 处未使用的 doc_ids 变量和 2 处过时注释

.dockerignore

@@ -0,0 +1,23 @@
+# 大体积数据目录（通过 volume 挂载，不需要打进镜像）
+models/
+knowledge/vector_store/
+documents/
+.data/
+data/
+
+# Python 虚拟环境
+venv/
+__pycache__/
+*.pyc
+
+# Git
+.git/
+
+# IDE 和编辑器
+.vscode/
+.idea/
+*.swp
+
+# 其他
+*.log
+.env*

api/document_routes.py

@@ -775,7 +775,22 @@ def delete_document(doc_path: str) -> Tuple[Any, int]:
         if kb_manager:
             kb_manager.delete_document(collection, filename)
 
-        # 2. 删除文件
+        # 2. 缓存失效
+        try:
+            from core.cache import get_cache_manager
+            _cm = get_cache_manager()
+            _cm.increment_kb_version(collection)
+        except Exception:
+            pass
+        try:
+            from core.semantic_cache import get_semantic_cache
+            _sc = get_semantic_cache()
+            if _sc:
+                _sc.clear()
+        except Exception:
+            pass
+
+        # 3. 删除文件
         os.remove(filepath)
 
         return jsonify({

api/kb_routes.py

@@ -260,6 +260,21 @@ def delete_collection(kb_name: str) -> Tuple[Any, int]:
     success, message = kb_manager.delete_collection(kb_name, delete_documents)
 
     if success:
+        # 缓存失效
+        try:
+            from core.cache import get_cache_manager
+            _cm = get_cache_manager()
+            _cm.increment_kb_version(kb_name)
+        except Exception:
+            pass
+        try:
+            from core.semantic_cache import get_semantic_cache
+            _sc = get_semantic_cache()
+            if _sc:
+                _sc.clear()
+        except Exception:
+            pass
+
         return jsonify({
             "success": True,
             "message": message,

cleanup_orphans.py

@@ -0,0 +1,207 @@
+"""
+孤儿文件清理工具
+
+清理 .data/images/ 和 .data/cache/vlm/ 中不再被任何 ChromaDB 切片引用的文件。
+
+用法:
+    # 预览模式（不删除，只显示孤儿文件）
+    python cleanup_orphans.py
+
+    # 实际删除
+    python cleanup_orphans.py --force
+
+    # 仅清理特定知识库
+    python cleanup_orphans.py --collections public_kb test_kb
+"""
+
+import argparse
+import hashlib
+import logging
+import os
+import sys
+from pathlib import Path
+
+logging.basicConfig(level=logging.INFO, format='%(levelname)s: %(message)s')
+logger = logging.getLogger(__name__)
+
+IMAGES_DIR = Path(".data/images")
+VLM_CACHE_DIR = Path(".data/cache/vlm")
+
+
+def compute_file_hash(file_path: str) -> str:
+    """计算文件 MD5"""
+    with open(file_path, 'rb') as f:
+        return hashlib.md5(f.read()).hexdigest()
+
+
+def collect_referenced_images(manager, collections=None) -> dict:
+    """
+    从 ChromaDB 收集所有被引用的图片路径。
+
+    Returns:
+        {image_filename: set of chunk_ids referencing it}
+    """
+    referenced = {}
+
+    if collections:
+        kb_names = collections
+    else:
+        kb_names = [c.name if hasattr(c, 'name') else str(c)
+                     for c in manager.list_collections()]
+
+    for kb_name in kb_names:
+        try:
+            col = manager.get_collection(kb_name)
+        except Exception as e:
+            logger.warning(f"无法获取 {kb_name}: {e}")
+            continue
+
+        # 查找所有带 image_path 的切片
+        results = col.get(include=['metadatas'])
+        if not results['ids']:
+            continue
+
+        for chunk_id, meta in zip(results['ids'], results['metadatas']):
+            image_path = meta.get('image_path', '')
+            if not image_path:
+                continue
+
+            # image_path 可能是: "185a7a75d246.png" 或相对路径
+            filename = os.path.basename(image_path)
+            if filename not in referenced:
+                referenced[filename] = set()
+            referenced[filename].add(f"{kb_name}/{chunk_id}")
+
+    return referenced
+
+
+def find_orphan_images(referenced: dict) -> list:
+    """
+    查找 .data/images/ 中不再被引用的图片文件。
+
+    Returns:
+        [(filepath, filename, size_bytes)]
+    """
+    orphans = []
+    if not IMAGES_DIR.exists():
+        return orphans
+
+    for f in IMAGES_DIR.iterdir():
+        if not f.is_file():
+            continue
+        if f.name not in referenced:
+            orphans.append((str(f), f.name, f.stat().st_size))
+
+    return orphans
+
+
+def find_orphan_vlm_caches(referenced: dict) -> list:
+    """
+    查找 .data/cache/vlm/ 中对应的图片已不存在的缓存文件。
+
+    缓存文件以图片 MD5 命名，如果图片被删了，缓存也应该是孤儿。
+
+    Returns:
+        [(filepath, filename, size_bytes)]
+    """
+    orphans = []
+    if not VLM_CACHE_DIR.exists():
+        return orphans
+
+    # 构建 referenced 中所有图片的 MD5 集合
+    referenced_hashes = set()
+    for filename in referenced:
+        full_path = IMAGES_DIR / filename
+        if full_path.exists():
+            try:
+                img_hash = compute_file_hash(str(full_path))
+                referenced_hashes.add(img_hash)
+            except Exception:
+                pass
+
+    for f in VLM_CACHE_DIR.iterdir():
+        if not f.is_file() or not f.suffix == '.txt':
+            continue
+        cache_hash = f.stem  # 文件名就是 MD5
+        if cache_hash not in referenced_hashes:
+            orphans.append((str(f), f.name, f.stat().st_size))
+
+    return orphans
+
+
+def delete_files(file_list: list, dry_run: bool = True) -> int:
+    """删除文件列表，返回删除数量"""
+    deleted = 0
+    for filepath, filename, size in file_list:
+        if dry_run:
+            logger.info(f"  [DRY-RUN] 将删除: {filename} ({size/1024:.1f} KB)")
+        else:
+            try:
+                os.remove(filepath)
+                logger.info(f"  已删除: {filename} ({size/1024:.1f} KB)")
+                deleted += 1
+            except OSError as e:
+                logger.warning(f"  删除失败: {filename} - {e}")
+    return deleted
+
+
+def main():
+    parser = argparse.ArgumentParser(description="孤儿文件清理工具")
+    parser.add_argument("--force", action="store_true",
+                        help="实际删除文件（默认仅预览）")
+    parser.add_argument("--collections", nargs="+",
+                        help="仅检查指定知识库（默认检查全部）")
+    args = parser.parse_args()
+
+    os.chdir(os.path.dirname(os.path.abspath(__file__)))
+
+    mode = "删除" if args.force else "预览（不删除）"
+    print("=" * 60)
+    print(f"孤儿文件清理 - {mode}")
+    print("=" * 60)
+
+    # 1. 收集引用
+    print("\n[1/4] 扫描 ChromaDB 引用...")
+    sys.path.insert(0, '.')
+    from knowledge.manager import get_kb_manager
+    manager = get_kb_manager()
+    referenced = collect_referenced_images(manager, args.collections)
+    print(f"  被引用的图片: {len(referenced)} 个")
+
+    # 2. 查找孤儿图片
+    print("\n[2/4] 查找孤儿图片...")
+    orphan_images = find_orphan_images(referenced)
+    total_img_size = sum(s for _, _, s in orphan_images)
+    print(f"  孤儿图片: {len(orphan_images)} 个 ({total_img_size/1024:.1f} KB)")
+
+    # 3. 查找孤儿 VLM 缓存
+    print("\n[3/4] 查找孤儿 VLM 缓存...")
+    orphan_caches = find_orphan_vlm_caches(referenced)
+    total_cache_size = sum(s for _, _, s in orphan_caches)
+    print(f"  孤儿缓存: {len(orphan_caches)} 个 ({total_cache_size/1024:.1f} KB)")
+
+    # 4. 清理
+    print("\n[4/4] 清理...")
+    if not orphan_images and not orphan_caches:
+        print("  没有需要清理的文件")
+    else:
+        if not args.force:
+            print("  预览模式，以下文件将被删除：")
+
+        img_deleted = delete_files(orphan_images, dry_run=not args.force)
+        cache_deleted = delete_files(orphan_caches, dry_run=not args.force)
+
+        if args.force:
+            print(f"\n  删除了 {img_deleted} 个图片 + {cache_deleted} 个缓存")
+            freed = total_img_size + total_cache_size
+            print(f"  释放空间: {freed/1024:.1f} KB")
+        else:
+            print(f"\n  共 {len(orphan_images) + len(orphan_caches)} 个文件待清理")
+            print(f"  释放空间: {(total_img_size + total_cache_size)/1024:.1f} KB")
+            print("  使用 --force 参数执行实际删除")
+
+    print("=" * 60)
+
+
+if __name__ == '__main__':
+    main()

config.example.py

@@ -109,6 +109,11 @@ USE_RERANK = True
 RERANK_CANDIDATES = 20           # 送入重排序的候选数
 RERANK_TOP_K = 15                # 重排序后保留数
 RERANK_USE_ONNX = os.getenv("RERANK_USE_ONNX", "true").lower() == "true"
+RERANK_BACKEND = os.getenv("RERANK_BACKEND", "local")       # local / cloud / fallback
+RERANK_CLOUD_MODEL = os.getenv("RERANK_CLOUD_MODEL", "qwen3-rerank")
+RERANK_CLOUD_API_KEY = os.getenv("RERANK_CLOUD_API_KEY", "")
+RERANK_CLOUD_BASE_URL = os.getenv("RERANK_CLOUD_BASE_URL", "https://dashscope.aliyuncs.com/compatible-api/v1/reranks")
+RERANK_CLOUD_TIMEOUT = int(os.getenv("RERANK_CLOUD_TIMEOUT", "15"))
 RERANK_CONTEXT_MIN_SCORE = 0.05  # Rerank 分数低于此值的切片不送入 LLM
 
 # ----- RRF 融合 -----

core/agentic.py

@@ -31,17 +31,11 @@ from core.llm_utils import call_llm, quick_yes_no, parse_json_from_response
 from .agentic_base import (
     API_KEY, BASE_URL, MODEL,
     HAS_SERPER,
-    HAS_BUDGET, SEMANTIC_CACHE_ENABLED,
+    HAS_BUDGET,
     MAX_CONTEXT_TOKENS, MAX_CONTEXT_COUNT, RERANK_THRESHOLD,
     SOURCE_KB, SOURCE_WEB,
 )
 
-# 尝试导入语义缓存
-try:
-    from core.semantic_cache import SemanticCache
-except ImportError:
-    SemanticCache = None
-
 # 导入 Mixin 类
 from .agentic_query import QueryRewriteMixin
 from .agentic_search import SearchMixin
@@ -120,30 +114,6 @@ class AgenticRAG(
         except ImportError:
             self.loop_guard = None
 
-        # 初始化语义缓存
-        self.semantic_cache = None
-        self.embedding_model = None
-        if SEMANTIC_CACHE_ENABLED and SemanticCache:
-            try:
-                engine = get_engine()
-                if engine and hasattr(engine, 'embedding_model'):
-                    self.embedding_model = engine.embedding_model
-                    emb_dim = 768
-                    # 优先使用新 API，兼容旧版本
-                    if hasattr(self.embedding_model, 'get_embedding_dimension'):
-                        emb_dim = self.embedding_model.get_embedding_dimension()
-                    elif hasattr(self.embedding_model, 'get_sentence_embedding_dimension'):
-                        emb_dim = self.embedding_model.get_sentence_embedding_dimension()
-                    self.semantic_cache = SemanticCache(
-                        dim=emb_dim,
-                        threshold=0.92,
-                        max_size=5000
-                    )
-                    logger.info(f"语义缓存已启用，维度={emb_dim}")
-            except Exception as e:
-                logger.warning(f"语义缓存初始化失败: {e}")
-                self.semantic_cache = None
-
         # Context Compression 配置
         self.MAX_CONTEXT_TOKENS = MAX_CONTEXT_TOKENS
         self.MAX_CONTEXT_COUNT = MAX_CONTEXT_COUNT

core/agentic_base.py

@@ -25,22 +25,6 @@ except ImportError:
     HAS_BUDGET = False
     CallType = None
 
-# 语义缓存
-try:
-    from config import SEMANTIC_CACHE_ENABLED, SEMANTIC_CACHE_THRESHOLD
-    HAS_SEMANTIC_CACHE_CONFIG = True
-except ImportError:
-    SEMANTIC_CACHE_ENABLED = False
-    SEMANTIC_CACHE_THRESHOLD = 0.92
-    HAS_SEMANTIC_CACHE_CONFIG = False
-
-try:
-    from core.semantic_cache import SemanticCache, get_semantic_cache
-    HAS_SEMANTIC_CACHE = True
-except ImportError:
-    HAS_SEMANTIC_CACHE = False
-    SemanticCache = None
-
 # LLM 配置
 try:
     from config import API_KEY, BASE_URL, MODEL

core/cache.py

@@ -189,6 +189,8 @@ class RAGCacheManager:
         # 失效旧版本缓存
         self.query_cache.invalidate_by_version(old_version)
         self.embedding_cache.invalidate_by_version(old_version)
+        # Rerank cache 无 kb_version 字段，文档变更时全量清空以防过时分数
+        self.rerank_cache.clear()
 
         logger.info(f"知识库 {kb_name} 版本更新: {old_version} -> {new_version}")
         return new_version
@@ -196,90 +198,41 @@ class RAGCacheManager:
     # ==================== Query Cache 方法 ====================
 
     @staticmethod
-    def _make_query_cache_key(query: str, kb_name: str, kb_version: int, doc_hash: str = "") -> str:
+    def _make_query_cache_key(query: str, kb_name: str, kb_version: int) -> str:
         """
-        生成查询缓存 key
-
-        Args:
-            query: 查询文本
-            kb_name: 知识库名称
-            kb_version: 知识库版本号
-            doc_hash: 相关文档版本哈希（细粒度失效）
-
-        Returns:
-            缓存 key
+        生成查询缓存 key（基于 kb_version 的粗粒度失效）
         """
-        if doc_hash:
-            # 细粒度：只失效相关文档的缓存
-            return hashlib.md5(
-                f"query:{query}:{kb_name}:{doc_hash}".encode()
-            ).hexdigest()
-        else:
-            # 粗粒度：整个知识库版本变化时失效
         return hashlib.md5(
             f"query:{query}:{kb_name}:{kb_version}".encode()
         ).hexdigest()
 
-    def get_query_result(self, query: str, kb_name: str, doc_ids: List[str] = None) -> Optional[Dict]:
+    def get_query_result(self, query: str, kb_name: str) -> Optional[Dict]:
         """
         获取查询缓存结果
 
         Args:
             query: 查询文本
             kb_name: 知识库名称
-            doc_ids: 相关文档 ID 列表（用于细粒度缓存 key）
         """
         kb_version = self.get_kb_version(kb_name)
-
-        # 计算文档哈希（如果提供了 doc_ids）
-        doc_hash = ""
-        if doc_ids:
-            doc_hash = self._compute_doc_hash(kb_name, doc_ids)
-
-        key = self._make_query_cache_key(query, kb_name, kb_version, doc_hash)
+        key = self._make_query_cache_key(query, kb_name, kb_version)
         return self.query_cache.get(key)
 
-    def set_query_result(self, query: str, kb_name: str, result: Dict, doc_ids: List[str] = None) -> None:
+    def set_query_result(self, query: str, kb_name: str, result: Dict) -> None:
         """
         设置查询缓存结果
 
+        kb_version 在文档变更时自增，触发整个知识库的缓存失效。
+
         Args:
             query: 查询文本
             kb_name: 知识库名称
             result: 缓存结果
-            doc_ids: 相关文档 ID 列表（用于细粒度失效）
         """
         kb_version = self.get_kb_version(kb_name)
-
-        # 计算相关文档的版本哈希（细粒度失效）
-        doc_hash = ""
-        if doc_ids:
-            doc_hash = self._compute_doc_hash(kb_name, doc_ids)
-
-        key = self._make_query_cache_key(query, kb_name, kb_version, doc_hash)
+        key = self._make_query_cache_key(query, kb_name, kb_version)
         self.query_cache.set(key, result, kb_version=kb_version)
 
-    def _compute_doc_hash(self, kb_name: str, doc_ids: List[str]) -> str:
-        """
-        计算文档版本哈希
-
-        用于细粒度缓存失效：只失效相关文档变化时的缓存
-        """
-        if not doc_ids:
-            return ""
-
-        # 从文档 ID 中提取 source（文件名）
-        sources = set()
-        for doc_id in doc_ids:
-            # doc_id 格式通常为 "filename_text_0" 或类似
-            parts = doc_id.split('_')
-            if parts:
-                sources.add(parts[0])
-
-        # 生成哈希
-        sources_str = ','.join(sorted(sources))
-        return hashlib.md5(f"docs:{sources_str}".encode()).hexdigest()
-
     # ==================== Embedding Cache 方法 ====================
 
     @staticmethod

core/engine.py

@@ -29,6 +29,7 @@ RAG 核心引擎
 
 import os
 import gc
+import re
 import time
 import logging
 import threading
@@ -76,6 +77,10 @@ try:
         CONTEXT_EXPANSION_MAX_CHUNKS, ENUM_QUERY_DISABLE_TOPK_SHRINK, ENUM_QUERY_MMR_LAMBDA,
         # Phase 3 扩展精细化
         EXPANSION_SCORE_THRESHOLD, MAX_EXPANDED_NEIGHBORS,
+        # 章节聚类救援
+        SECTION_CLUSTER_BOOST_ENABLED, CLUSTER_MIN_MEMBERS, CLUSTER_MIN_TYPES,
+        CLUSTER_SEED_FLOOR, CLUSTER_MAX_BOOST_PER_SECTION, CLUSTER_MAX_SECTIONS,
+        CLUSTER_SECTION_PREFIX_LEVELS,
         # 上下文与生成
         LLM_TEMPERATURE, LLM_MAX_TOKENS, RECALL_MULTIPLIER,
         # FAQ 与黑名单
@@ -101,20 +106,28 @@ except ImportError:
     MMR_TOP_K = 30
     CONTEXT_EXPANSION_ENABLED = True
     CONTEXT_EXPANSION_BEFORE = 1
-    CONTEXT_EXPANSION_AFTER = 5
+    CONTEXT_EXPANSION_AFTER = 8
     CONTEXT_EXPANSION_MAX_CHUNKS = 24
     EXPANSION_SCORE_THRESHOLD = 0.3
-    MAX_EXPANDED_NEIGHBORS = 4
+    MAX_EXPANDED_NEIGHBORS = 8
+    # 章节聚类救援默认值
+    SECTION_CLUSTER_BOOST_ENABLED = True
+    CLUSTER_MIN_MEMBERS = 3
+    CLUSTER_MIN_TYPES = 2
+    CLUSTER_SEED_FLOOR = 0.35
+    CLUSTER_MAX_BOOST_PER_SECTION = 8
+    CLUSTER_MAX_SECTIONS = 3
+    CLUSTER_SECTION_PREFIX_LEVELS = 1
     ENUM_QUERY_DISABLE_TOPK_SHRINK = True
     ENUM_QUERY_MMR_LAMBDA = 0.85
     DYNAMIC_RRF_ENABLED = True
     EMBEDDING_DEVICE = "auto"
     RERANK_DEVICE = "auto"
     RERANK_USE_ONNX = False
-    RERANK_BACKEND = "local"
-    RERANK_CLOUD_MODEL = "qwen3-rerank"
+    RERANK_BACKEND = "cloud"
+    RERANK_CLOUD_MODEL = "xop3qwen8breranker"
     RERANK_CLOUD_API_KEY = ""
-    RERANK_CLOUD_BASE_URL = "https://dashscope.aliyuncs.com/compatible-api/v1/reranks"
+    RERANK_CLOUD_BASE_URL = "https://maas-api.cn-huabei-1.xf-yun.com/v2/rerank"
     RERANK_CLOUD_TIMEOUT = 15
 
 
@@ -272,8 +285,7 @@ class CloudReranker:
         body = {
             "model": self.model,
             "query": query,
-            "documents": documents,
-            "top_n": len(documents)
+            "documents": documents
         }
 
         session = self._get_session()
@@ -547,8 +559,7 @@ class RAGEngine:
                 top_dist = result['distances'][0][0] if result.get('distances') and result['distances'][0] else 1.0
                 top_score = 1.0 - top_dist
                 if top_score >= CACHE_MIN_SCORE:
-                    doc_ids = result.get('ids', [[]])[0] if result.get('ids') else []
-                    cache.set_query_result(query, kb_name, result, doc_ids=doc_ids)
+                    cache.set_query_result(query, kb_name, result)
             result['_debug'] = _debug
             return result
 
@@ -569,8 +580,7 @@ class RAGEngine:
                         top_dist = result['distances'][0][0] if result.get('distances') and result['distances'][0] else 1.0
                         top_score = 1.0 - top_dist
                         if top_score >= CACHE_MIN_SCORE:
-                            doc_ids = result.get('ids', [[]])[0] if result.get('ids') else []
-                            cache.set_query_result(query, kb_name, result, doc_ids=doc_ids)
+                            cache.set_query_result(query, kb_name, result)
                     result['_debug'] = _debug
                     return result
             except Exception as e:
@@ -605,9 +615,7 @@ class RAGEngine:
                 top_dist = result['distances'][0][0] if result.get('distances') and result['distances'][0] else 1.0
                 top_score = 1.0 - top_dist
                 if top_score >= CACHE_MIN_SCORE:
-                    # 传递 doc_ids 实现细粒度缓存失效
-                    doc_ids = result.get('ids', [[]])[0] if result.get('ids') else []
-                    cache.set_query_result(query, kb_name, result, doc_ids=doc_ids)
+                    cache.set_query_result(query, kb_name, result)
             _debug['timing']['total_ms'] = int((time.time() - _overall_start) * 1000)
             result['_debug'] = _debug
             return result
@@ -625,7 +633,7 @@ 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)
 
@@ -657,6 +665,7 @@ class RAGEngine:
 
         results_list = [vector_results]
         weights = [VECTOR_WEIGHT]
+        bm25_results = None  # 初始化，防止 NameError
 
         if USE_HYBRID_SEARCH and self.bm25_index.bm25:
             bm25_results = self.bm25_index.search(query, top_k=recall_k)
@@ -671,6 +680,22 @@ class RAGEngine:
             vector_w, bm25_w = self._get_dynamic_rrf_weights(query)
             weights = [vector_w, bm25_w]
 
+        # ========== 保留 BM25 原始 top-3 完整信息，用于下游分歧检测救援 ==========
+        _bm25_raw_top3 = []
+        if USE_HYBRID_SEARCH and bm25_results and bm25_results.get('ids') and bm25_results['ids'][0]:
+            _bm25_ids = bm25_results['ids'][0][:3]
+            _bm25_docs = bm25_results['documents'][0][:3]
+            _bm25_metas = bm25_results['metadatas'][0][:3]
+            _bm25_dists = (bm25_results.get('distances', [[]])[0] or [0]*3)[:3]
+            for i in range(len(_bm25_ids)):
+                _bm25_raw_top3.append({
+                    'id': _bm25_ids[i],
+                    'doc': _bm25_docs[i],
+                    'meta': _bm25_metas[i],
+                    'bm25_score': _bm25_dists[i],
+                    'rank': i + 1
+                })
+
         if len(results_list) > 1:
             fused_results = self.reciprocal_rank_fusion(results_list, weights)
             _debug['steps'].append({'name': 'rrf_fusion', 'count': len(fused_results['ids'][0]) if fused_results.get('ids') else 0, 'weights': [round(w, 2) for w in weights]})
@@ -686,6 +711,8 @@ class RAGEngine:
 
         is_enum_query = self._is_enumeration_query(query)
         fused_results['_enum_query'] = is_enum_query
+        # 传递 BM25 原始 top-3 到路由层，用于分歧检测救援
+        fused_results['_bm25_top3'] = _bm25_raw_top3
 
         # 章节过滤（如果查询中提到了章节）
         fused_results = self._filter_by_section(fused_results, query)
@@ -732,11 +759,19 @@ class RAGEngine:
         # 时间衰减（Time Decay）
         fused_results = self._apply_time_decay(fused_results)
 
+        # 提前附加 _debug，使聚类提升能写入调试步骤
+        fused_results['_debug'] = _debug
+
+        # ========== 章节聚类提升：在扩展前将低分但聚类的切片提升至种子阈值 ==========
+        if SECTION_CLUSTER_BOOST_ENABLED:
+            fused_results = self._section_cluster_boost(fused_results, query)
+
         # ========== 上下文扩展：补充强命中切片周围的连续文本（rerank 之后，防止被截断）==========
         # Phase 3：仅对高分种子扩展邻居
         before_exp = len(fused_results['ids'][0]) if fused_results.get('ids') else 0
         fused_results = self._expand_contiguous_chunks(fused_results, top_k=top_k,
-                                                        min_score=EXPANSION_SCORE_THRESHOLD)
+                                                        min_score=EXPANSION_SCORE_THRESHOLD,
+                                                        query=query)
         after_exp = len(fused_results['ids'][0]) if fused_results.get('ids') else 0
         _debug['steps'].append({'name': 'context_expansion', 'before': before_exp, 'after': after_exp})
 
@@ -748,7 +783,15 @@ class RAGEngine:
             and not (is_enum_query and ENUM_QUERY_DISABLE_TOPK_SHRINK)
             and fused_results.get('_score_source') != 'rrf'
         ):
-            top_score = 1.0 - fused_results['distances'][0][0]  # 距离转相似度
+            # 根据分数来源计算相似度分数（越高越好）
+            score_source = fused_results.get('_score_source')
+            top_dist = fused_results['distances'][0][0]
+            if score_source == 'rerank':
+                # Rerank 后 distances 是相关性分数，越大越好，直接使用
+                top_score = top_dist
+            else:
+                # 向量距离，越小越好，转为相似度
+                top_score = 1.0 - top_dist
             adjusted_k, should_retrieve, reason = self._adaptive_topk.adjust(top_score, top_k)
             if "high_confidence" in reason:
                 # 高置信度时截断结果
@@ -763,9 +806,7 @@ class RAGEngine:
             top_dist = fused_results['distances'][0][0] if fused_results.get('distances') and fused_results['distances'][0] else 1.0
             top_score = 1.0 - top_dist  # 距离转相似度
             if top_score >= CACHE_MIN_SCORE:  # 置信度阈值
-                # 传递 doc_ids 实现细粒度缓存失效
-                doc_ids = fused_results.get('ids', [[]])[0] if fused_results.get('ids') else []
-                cache.set_query_result(query, kb_name, fused_results, doc_ids=doc_ids)
+                cache.set_query_result(query, kb_name, fused_results)
 
         fused_results['_debug'] = _debug
         _debug['timing']['total_ms'] = int((time.time() - _overall_start) * 1000)
@@ -811,6 +852,76 @@ 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）
@@ -1017,7 +1128,7 @@ class RAGEngine:
             'metadatas': [f_metas],
             'distances': [f_scores]
         }
-        for key in ('_debug', '_score_source', '_enum_query', '_expanded_context'):
+        for key in ('_debug', '_score_source', '_enum_query', '_expanded_context', '_bm25_top3'):
             if key in results:
                 filtered[key] = results[key]
         return filtered
@@ -1040,7 +1151,7 @@ class RAGEngine:
             'metadatas': [f_metas],
             'distances': [f_scores]
         }
-        for key in ('_debug', '_score_source', '_enum_query', '_expanded_context'):
+        for key in ('_debug', '_score_source', '_enum_query', '_expanded_context', '_bm25_top3'):
             if key in results:
                 filtered[key] = results[key]
         return filtered
@@ -1055,7 +1166,7 @@ class RAGEngine:
             'metadatas': [results['metadatas'][0][:top_k]],
             'distances': [results['distances'][0][:top_k]]
         }
-        for key in ('_debug', '_score_source', '_enum_query', '_expanded_context'):
+        for key in ('_debug', '_score_source', '_enum_query', '_expanded_context', '_bm25_top3'):
             if key in results:
                 truncated[key] = results[key]
         return truncated
@@ -1090,14 +1201,139 @@ class RAGEngine:
                 return None
         return self.collection
 
+    @staticmethod
+    def _normalize_section_path(section_path: str, levels: int = None) -> str:
+        """归一化 section_path：取前 N 级路径，容忍 MinerU 标题检测误差。
+
+        例如: "第三章 吸烟场所的功能设置 > 第三条 文明吸烟..." → "第三章 吸烟场所的功能设置"
+        """
+        if not section_path:
+            return ''
+        if levels is None:
+            levels = CLUSTER_SECTION_PREFIX_LEVELS
+        parts = [p.strip() for p in section_path.split('>')]
+        return ' > '.join(parts[:levels])
+
+    def _section_cluster_boost(self, results: dict, query: str = '') -> dict:
+        """章节聚类提升：当同一 section 下多个切片（text+table）同时出现在候选集中，
+        即使单个切片 CrossEncoder 分数很低，也将整组提升到种子阈值。
+
+        核心洞察：单个低分切片不可信，但同一 section 多个切片同时出现是强信号。
+        提升后的切片可以作为 _expand_contiguous_chunks 的种子，触发邻居扩展。
+
+        Args:
+            results: rerank 后的检索结果
+            query: 用户查询（用于后续扩展）
+
+        Returns:
+            修改后的 results（distances 被调整，meta 中标记 _cluster_boosted）
+        """
+        if not results.get('ids') or not results['ids'][0]:
+            return results
+
+        ids = results['ids'][0]
+        metas = results.get('metadatas', [[]])[0]
+        distances = results.get('distances', [[]])[0] if results.get('distances') else None
+
+        if not distances:
+            return results
+
+        # 1. 按 (source, normalized_section) 分组
+        from collections import defaultdict
+        section_groups = defaultdict(list)  # key → [(index, meta, dist)]
+
+        for i, (meta, dist) in enumerate(zip(metas, distances)):
+            source = meta.get('source', '')
+            section_path = meta.get('section', '') or meta.get('section_path', '')
+            norm_section = self._normalize_section_path(section_path)
+            if not source or not norm_section:
+                continue
+            key = (source, norm_section)
+            section_groups[key].append((i, meta, dist))
+
+        # 2. 检测聚类信号并提升
+        boost_target_dist = 1.0 - CLUSTER_SEED_FLOOR  # score=0.35 → dist=0.65
+        boosted_sections = []
+        total_boosted = 0
+
+        # 按组成员数降序排列，优先处理最大聚类
+        sorted_groups = sorted(section_groups.items(), key=lambda x: len(x[1]), reverse=True)
+
+        for (source, norm_section), members in sorted_groups:
+            if len(boosted_sections) >= CLUSTER_MAX_SECTIONS:
+                break
+
+            # 聚类信号检测：成员数 >= 阈值 且 类型多样性 >= 阈值
+            chunk_types = set(m[1].get('chunk_type', 'text') for m in members)
+            if len(members) < CLUSTER_MIN_MEMBERS or len(chunk_types) < CLUSTER_MIN_TYPES:
+                continue
+
+            # 提升组内切片分数（仅提升低于阈值的）
+            boost_count = 0
+            for idx, meta, dist in members:
+                if boost_count >= CLUSTER_MAX_BOOST_PER_SECTION:
+                    break
+                # 只提升分数低于种子阈值的切片（高分切片不需要）
+                if dist > boost_target_dist:
+                    distances[idx] = boost_target_dist
+                    meta['_cluster_boosted'] = True
+                    boost_count += 1
+                    total_boosted += 1
+
+            if boost_count > 0:
+                boosted_sections.append({
+                    'source': source,
+                    'section': norm_section,
+                    'members': len(members),
+                    'types': list(chunk_types),
+                    'boosted': boost_count
+                })
+
+        # 3. 写 debug 信息
+        if boosted_sections:
+            debug_info = results.get('_debug', {})
+            if 'steps' not in debug_info:
+                debug_info['steps'] = []
+            debug_info['steps'].append({
+                'name': 'section_cluster_boost',
+                'sections': boosted_sections,
+                'total_boosted': total_boosted
+            })
+            results['_debug'] = debug_info
+            logger.info(f"[章节聚类提升] 提升 {total_boosted} 个切片，"
+                       f"涉及 {len(boosted_sections)} 个 section: "
+                       f"{[s['section'][:30] for s in boosted_sections]}")
+
+        return results
+
+    @staticmethod
+    def _chunk_lexical_score(chunk_text: str, query: str) -> float:
+        """计算切片文本与查询的词法重叠度（bigram 命中率），用于辅助种子资格判定。"""
+        if not chunk_text or not query:
+            return 0.0
+        import re
+        clean_q = re.sub(r'[？?！!。，,、；;：:"""\'\s*#`]+', ' ', query).strip()
+        if len(clean_q) < 2:
+            return 0.0
+        bigrams = set()
+        for i in range(len(clean_q) - 1):
+            w = clean_q[i:i+2].strip()
+            if len(w) == 2:
+                bigrams.add(w)
+        if not bigrams:
+            return 0.0
+        matched = sum(1 for w in bigrams if w in chunk_text)
+        return matched / len(bigrams)
+
     def _expand_contiguous_chunks(self, results: dict, top_k: int = None,
-                                   min_score: float = 0.0) -> dict:
+                                   min_score: float = 0.0, query: str = '') -> dict:
         """Add same-source same-section neighbor text chunks around strong hits.
 
         Args:
             results: 检索结果
             top_k: 最大切片数
             min_score: Phase 3 最低分数阈值，仅对 Rerank 分数高于此值的种子扩展
+            query: 查询文本，用于词法匹配辅助种子资格判定
         """
         if not CONTEXT_EXPANSION_ENABLED:
             return results
@@ -1127,7 +1363,7 @@ class RAGEngine:
         seeds = [
             (doc_id, doc, meta, dist)
             for doc_id, doc, meta, dist in items[:base_limit]
-            if meta.get('chunk_type', 'text') == 'text'
+            if (meta.get('chunk_type', 'text') == 'text' or meta.get('_cluster_boosted'))
             and meta.get('source')
             and self._to_int(meta.get('chunk_index')) is not None
         ]
@@ -1138,7 +1374,11 @@ class RAGEngine:
                 break
 
             # Phase 3：跳过分数低于阈值的种子（仅当 min_score > 0 时生效）
+            # 词法匹配豁免：CrossEncoder 低分但关键词重叠度高时仍允许作为种子
             if min_score > 0 and seed_dist < min_score:
+                if query and self._chunk_lexical_score(_seed_doc, query) > 0.3:
+                    pass  # 词法匹配度高，允许作为种子
+                else:
                     continue
 
             source = seed_meta.get('source')
@@ -1161,6 +1401,7 @@ class RAGEngine:
                     logger.warning(f"扩展连续切片失败: {e}")
                     return {'ids': [], 'documents': [], 'metadatas': []}
 
+            # 扩展同 section 的 text 邻居
             where_filter = {"$and": [{"source": source}, {"chunk_type": "text"}]}
             if section:
                 where_filter["$and"].append({"section": section})
@@ -1171,6 +1412,20 @@ class RAGEngine:
             if not neighbors.get('ids') or len(neighbors.get('ids', [])) <= 1:
                 neighbors = _get_neighbors({"$and": [{"source": source}, {"chunk_type": "text"}]})
 
+            # 同时扩展同 section 的 table 邻居（table 切片的 rerank 分数往往偏低，
+            # 但与同 section 的 text 切片属于同一语义单元，不应割裂）
+            table_where = {"$and": [{"source": source}, {"chunk_type": "table"}]}
+            if section:
+                table_where["$and"].append({"section": section})
+            table_neighbors = _get_neighbors(table_where)
+
+            # 当 section 为空时，table 查询只有 source 条件，可能拉入大量无关表格，
+            # 缩小 chunk_index 窗口至 ±1 以降低噪音；有 section 时使用正常窗口
+            if section:
+                _t_before, _t_after = CONTEXT_EXPANSION_BEFORE, CONTEXT_EXPANSION_AFTER
+            else:
+                _t_before, _t_after = 1, 1
+
             neighbor_rows = []
             for n_id, n_doc, n_meta in zip(
                 neighbors.get('ids', []),
@@ -1183,6 +1438,18 @@ class RAGEngine:
                 if seed_index - CONTEXT_EXPANSION_BEFORE <= n_index <= seed_index + CONTEXT_EXPANSION_AFTER:
                     neighbor_rows.append((n_index, n_id, n_doc, n_meta))
 
+            # 同 section 的 table 邻居也加入扩展范围
+            for n_id, n_doc, n_meta in zip(
+                table_neighbors.get('ids', []),
+                table_neighbors.get('documents', []),
+                table_neighbors.get('metadatas', [])
+            ):
+                n_index = self._to_int(n_meta.get('chunk_index'))
+                if n_index is None:
+                    continue
+                if seed_index - _t_before <= n_index <= seed_index + _t_after:
+                    neighbor_rows.append((n_index, n_id, n_doc, n_meta))
+
             seed_neighbors_added = 0
             for n_index, n_id, n_doc, n_meta in sorted(neighbor_rows, key=lambda row: row[0]):
                 if len(items) >= max_chunks:
@@ -1214,7 +1481,7 @@ class RAGEngine:
             'distances': [[item[3] for item in items]],
             '_expanded_context': {'added': added}
         }
-        for key in ('_debug', '_score_source', '_enum_query'):
+        for key in ('_debug', '_score_source', '_enum_query', '_bm25_top3'):
             if key in results:
                 expanded[key] = results[key]
         return expanded
@@ -1241,6 +1508,7 @@ class RAGEngine:
         sub_top_k = max(top_k, 5)
 
         all_results = []
+        _all_bm25_top3 = []  # 收集各子查询的 BM25 top3
         for sub_q in sub_queries:
             try:
                 sub_result = self.search_knowledge(
@@ -1251,6 +1519,9 @@ class RAGEngine:
                 )
                 if sub_result and sub_result.get('ids') and sub_result['ids'][0]:
                     all_results.append(sub_result)
+                    # 收集子查询的 BM25 top3
+                    if sub_result.get('_bm25_top3'):
+                        _all_bm25_top3.extend(sub_result['_bm25_top3'])
             except Exception as e:
                 logger.warning(f"子查询检索失败: '{sub_q}' - {e}")
 
@@ -1259,9 +1530,19 @@ class RAGEngine:
 
         # 合并去重
         if len(all_results) == 1:
-            return all_results[0]
+            merged = all_results[0]
+        else:
+            merged = self._merge_and_deduplicate(all_results, top_k)
 
-        return self._merge_and_deduplicate(all_results, top_k)
+        # 将收集的 BM25 top3 传递到合并结果中
+        if _all_bm25_top3:
+            _all_bm25_top3.sort(key=lambda x: x.get('bm25_score', 0), reverse=True)
+            _all_bm25_top3 = _all_bm25_top3[:3]
+            for rank, item in enumerate(_all_bm25_top3):
+                item['rank'] = rank + 1
+            merged['_bm25_top3'] = _all_bm25_top3
+
+        return merged
 
     def _search_with_decomposition(
         self, query, decomposer, top_k=5, allowed_levels=None,
@@ -1293,6 +1574,7 @@ class RAGEngine:
 
         # 并行检索各子查询
         all_results = []
+        _all_bm25_top3 = []  # 收集各子查询的 BM25 top3
         for sub_q in sub_queries:
             try:
                 sub_result = self.search_knowledge(
@@ -1303,6 +1585,9 @@ class RAGEngine:
                 )
                 if sub_result and sub_result.get('ids') and sub_result['ids'][0]:
                     all_results.append(sub_result)
+                    # 收集子查询的 BM25 top3
+                    if sub_result.get('_bm25_top3'):
+                        _all_bm25_top3.extend(sub_result['_bm25_top3'])
             except Exception as e:
                 logger.warning(f"子查询检索失败: '{sub_q}' - {e}")
 
@@ -1315,6 +1600,14 @@ class RAGEngine:
         else:
             merged = self._merge_and_deduplicate(all_results, top_k)
 
+        # 将收集的 BM25 top3 传递到合并结果中
+        if _all_bm25_top3:
+            _all_bm25_top3.sort(key=lambda x: x.get('bm25_score', 0), reverse=True)
+            _all_bm25_top3 = _all_bm25_top3[:3]
+            for rank, item in enumerate(_all_bm25_top3):
+                item['rank'] = rank + 1
+            merged['_bm25_top3'] = _all_bm25_top3
+
         return merged
 
     def _merge_and_deduplicate(self, results_list, top_k):
@@ -1387,7 +1680,7 @@ 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)
@@ -1399,12 +1692,13 @@ class RAGEngine:
         from concurrent.futures import ThreadPoolExecutor, as_completed
 
         def _query_single_collection(coll_name):
-            """查询单个向量库（向量 + BM25）"""
+            """查询单个向量库（向量 + BM25），返回 (coll_results, bm25_raw_items)"""
             coll_results = []
+            bm25_raw_items = []  # 该 collection 的 BM25 原始结果
             try:
                 coll = self.kb_manager.get_collection(coll_name)
                 if not coll:
-                    return coll_results
+                    return coll_results, bm25_raw_items
 
                 query_kwargs = {
                     "query_embeddings": [query_vector],
@@ -1422,25 +1716,61 @@ class RAGEngine:
                 if USE_HYBRID_SEARCH:
                     try:
                         bm25 = self.kb_manager.get_bm25_index(coll_name)
-                        if bm25.bm25:
+                        if bm25 and bm25.bm25:
                             bm25_res = bm25.search(query, top_k=recall_k)
-                            if source_filter and bm25_res['metadatas'] and bm25_res['metadatas'][0]:
+                            # 兼容两种 BM25Index：core.bm25_index 返回 dict，knowledge.base 返回 tuple
+                            if isinstance(bm25_res, tuple):
+                                _ids, _docs, _metas, _dists = bm25_res
+                                bm25_res = {
+                                    'ids': [_ids],
+                                    'documents': [_docs],
+                                    'metadatas': [_metas],
+                                    'distances': [_dists]
+                                }
+                            if source_filter and bm25_res['metadatas'][0]:
                                 bm25_res = self._filter_results(bm25_res, lambda meta: meta.get('source') == source_filter)
                             if bm25_res['metadatas'] and bm25_res['metadatas'][0]:
                                 for meta in bm25_res['metadatas'][0]:
                                     meta['_collection'] = coll_name
                                 coll_results.append(bm25_res)
+                                # 提取 BM25 原始 top-3（在此处直接捕获，避免与向量结果混淆）
+                                _bm25_ids = bm25_res['ids'][0][:3]
+                                _bm25_docs = bm25_res['documents'][0][:3]
+                                _bm25_metas = bm25_res['metadatas'][0][:3]
+                                _bm25_dists = (bm25_res.get('distances', [[]])[0] or [0]*3)[:3]
+                                for i in range(len(_bm25_ids)):
+                                    # 确保 meta 包含 _collection（用于路由层注入时下游处理）
+                                    bm25_meta = _bm25_metas[i]
+                                    if '_collection' not in bm25_meta:
+                                        bm25_meta = {**bm25_meta, '_collection': coll_name}
+                                    bm25_raw_items.append({
+                                        'id': _bm25_ids[i],
+                                        'doc': _bm25_docs[i],
+                                        'meta': bm25_meta,
+                                        'bm25_score': _bm25_dists[i],
+                                    })
+                                logger.debug(f"[BM25] {coll_name}: captured {len(bm25_raw_items)} raw items")
                     except Exception as e:
-                        logger.debug(f"向量库 {coll_name} 检索失败: {e}")
+                        logger.debug(f"向量库 {coll_name} BM25检索失败: {e}")
             except Exception as e:
                 logger.debug(f"多向量库检索失败: {e}")
-            return coll_results
+            return coll_results, bm25_raw_items
 
         all_results = []
+        _bm25_raw_top3 = []
         with ThreadPoolExecutor(max_workers=len(target_collections)) as executor:
             futures = {executor.submit(_query_single_collection, name): name for name in target_collections}
             for future in as_completed(futures):
-                all_results.extend(future.result())
+                coll_results, bm25_raw_items = future.result()
+                all_results.extend(coll_results)
+                _bm25_raw_top3.extend(bm25_raw_items)
+
+        # 按 bm25_score 降序取全局 top-3
+        if _bm25_raw_top3:
+            _bm25_raw_top3.sort(key=lambda x: x['bm25_score'], reverse=True)
+            _bm25_raw_top3 = _bm25_raw_top3[:3]
+            for rank, item in enumerate(_bm25_raw_top3):
+                item['rank'] = rank + 1
 
         # ========== FAQ 检索 ==========
         faq_results = self._search_faq_collection(query_vector, top_k=FAQ_RECALL_TOP_K)
@@ -1488,6 +1818,8 @@ class RAGEngine:
 
         is_enum_query = self._is_enumeration_query(query)
         fused_results['_enum_query'] = is_enum_query
+        # 传递 BM25 原始 top-3 到路由层，用于分歧检测救援
+        fused_results['_bm25_top3'] = _bm25_raw_top3
 
         # 章节过滤（如果查询中提到了章节）
         fused_results = self._filter_by_section(fused_results, query)
@@ -1530,11 +1862,19 @@ class RAGEngine:
         # 时间衰减
         fused_results = self._apply_time_decay(fused_results)
 
+        # 提前附加 _debug，使聚类提升能写入调试步骤
+        fused_results['_debug'] = _debug
+
+        # ========== 章节聚类提升：在扩展前将低分但聚类的切片提升至种子阈值 ==========
+        if SECTION_CLUSTER_BOOST_ENABLED:
+            fused_results = self._section_cluster_boost(fused_results, query)
+
         # ========== 上下文扩展：补充强命中切片周围的连续文本（rerank 之后，防止被截断）==========
         # Phase 3：仅对高分种子扩展邻居
         before_exp = len(fused_results['ids'][0]) if fused_results.get('ids') else 0
         fused_results = self._expand_contiguous_chunks(fused_results, top_k=top_k,
-                                                        min_score=EXPANSION_SCORE_THRESHOLD)
+                                                        min_score=EXPANSION_SCORE_THRESHOLD,
+                                                        query=query)
         after_exp = len(fused_results['ids'][0]) if fused_results.get('ids') else 0
         if _debug is not None:
             _debug['steps'].append({'name': 'context_expansion', 'before': before_exp, 'after': after_exp})
@@ -1547,7 +1887,15 @@ class RAGEngine:
             and not (is_enum_query and ENUM_QUERY_DISABLE_TOPK_SHRINK)
             and fused_results.get('_score_source') != 'rrf'
         ):
-            top_score = 1.0 - fused_results['distances'][0][0]  # 距离转相似度
+            # 根据分数来源计算相似度分数（越高越好）
+            score_source = fused_results.get('_score_source')
+            top_dist = fused_results['distances'][0][0]
+            if score_source == 'rerank':
+                # Rerank 后 distances 是相关性分数，越大越好，直接使用
+                top_score = top_dist
+            else:
+                # 向量距离，越小越好，转为相似度
+                top_score = 1.0 - top_dist
             adjusted_k, should_retrieve, reason = self._adaptive_topk.adjust(top_score, top_k)
             if "high_confidence" in reason:
                 # 高置信度时截断结果
@@ -1597,7 +1945,7 @@ class RAGEngine:
             'metadatas': [filtered_metas],
             'distances': [filtered_distances]
         }
-        for key in ('_debug', '_score_source', '_enum_query', '_expanded_context'):
+        for key in ('_debug', '_score_source', '_enum_query', '_expanded_context', '_bm25_top3'):
             if key in results:
                 filtered[key] = results[key]
         return filtered
@@ -1670,7 +2018,7 @@ class RAGEngine:
             'metadatas': [filtered_metas],
             'distances': [filtered_distances]
         }
-        for key in ('_debug', '_score_source', '_enum_query', '_expanded_context'):
+        for key in ('_debug', '_score_source', '_enum_query', '_expanded_context', '_bm25_top3'):
             if key in results:
                 filtered[key] = results[key]
         return filtered
@@ -1739,12 +2087,12 @@ 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 = []
@@ -1786,7 +2134,7 @@ class RAGEngine:
                     'metadatas': [[c['metadata'] for c in selected]],
                     'distances': [[id_to_dist.get(doc_id, 0) for doc_id in selected_ids]]
                 }
-                for key in ('_debug', '_score_source', '_enum_query', '_expanded_context'):
+                for key in ('_debug', '_score_source', '_enum_query', '_expanded_context', '_bm25_top3'):
                     if key in results:
                         filtered[key] = results[key]
                 return filtered
@@ -1826,7 +2174,7 @@ class RAGEngine:
                     'metadatas': [[c['metadata'] for c in selected]],
                     'distances': [[id_to_dist.get(c['id'], 0) for c in selected]]
                 }
-                for key in ('_debug', '_score_source', '_enum_query', '_expanded_context'):
+                for key in ('_debug', '_score_source', '_enum_query', '_expanded_context', '_bm25_top3'):
                     if key in results:
                         filtered[key] = results[key]
                 return filtered
@@ -1935,109 +2283,15 @@ class RAGEngine:
             '_rerank_cached': cache_hit
         }
         # 保留原有标记字段
-        for key in ('_debug', '_score_source', '_enum_query', '_expanded_context'):
+        for key in ('_debug', '_enum_query', '_expanded_context', '_bm25_top3'):
             if key in results:
                 reranked[key] = results[key]
+        # Rerank 后 distances 语义变为 CrossEncoder 分数，更新 _score_source
+        # 使自适应 TopK 能正确应用（之前 _score_source='rrf' 会导致自适应 TopK 被跳过）
+        reranked['_score_source'] = 'rerank'
         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):
         """
@@ -2068,21 +2322,33 @@ class RAGEngine:
             "content": (
                 "你是一个严谨的知识库问答助手。"
                 "你必须且只能根据用户提供的【参考资料】回答问题。"
+                "参考资料中每段内容前标有章节路径（━格式），请注意区分不同章节的内容，"
+                "特别当不同章节标题相似或包含相同关键词时，务必根据章节路径准确定位，不要混淆。"
                 "如果参考资料中有答案，必须引用对应内容回答，并在回答末尾标注引用编号（如[1]、[2]）。"
                 "如果参考资料中确实没有相关信息，简短说明即可，不要编造或补充资料外的内容。"
                 "禁止使用参考资料以外的知识进行补充或推测。"
+                "【重要-表格处理规则】当用户询问表格、要求展示表格内容时，你必须将参考资料中的 Markdown 表格原样输出（保留 | 分隔符和表格结构），"
+                "不要仅用文字描述表格存在或仅列出章节名称。如果参考资料中多个章节都有表格，"
+                "优先展示与用户问题最相关的表格完整内容。"
             )
         })
 
         # 添加当前问题（带上下文）- 强化指令
         if context:
+            # 检测用户问题是否涉及表格，加入针对性指令
+            _table_hint = ""
+            # 检测上下文中是否包含 Markdown 表格（数据驱动，无需硬编码关键词）
+            _has_table_in_context = bool(re.search(r'\|.+\|', context)) if context else False
+            if _has_table_in_context:
+                _table_hint = "\n注意：参考资料中包含 Markdown 格式的表格数据，请务必将相关表格以原始 Markdown 表格格式完整展示在回答中，不要仅用文字描述。"
+
             user_message = f"""【参考资料】
 {context}
 
 【用户问题】
 {query}
 
-请仔细阅读以上全部参考资料后回答。如果参考资料中包含相关内容，必须引用回答并标注编号。如果资料中没有相关信息，请明确说明。"""
+请仔细阅读以上全部参考资料后回答。注意参考资料中标有章节路径，请根据章节路径准确定位相关内容。如果参考资料中包含相关内容，必须引用回答并标注编号。如果资料中没有相关信息，请明确说明。{_table_hint}"""
         else:
             user_message = query
 

core/intent_analyzer.py

@@ -83,9 +83,16 @@ class IntentAnalyzer:
 根据对话历史和当前用户消息，输出一个 JSON 对象，包含以下字段：
 
 1. **rewritten_query**: 改写后的完整问题
-   - 如果问题包含指代（如"这两张图片"、"继续说"），将其改写为完整、独立的问题
+   - **指代消解**：如果问题包含指代（如"这两张图片"、"继续说"），将其改写为完整、独立的问题
      - 例如："分析一下这两张图片" → "分析一下对话历史中提到的图片"
-   - 如果问题本身已经完整，直接返回原文
+   - **追问补全**：如果问题是省略式追问（省略了上一轮讨论的主题实体），必须补全为完整问题
+     - 判断方法：当前问题缺少主语/宾语，且对话历史中可以推断出省略的实体
+     - 补全方法：从上一轮用户问题中提取主题实体，与追问组合成完整问题
+     - 例如：
+       - 上一轮问"吸烟点C1类是什么区？"，追问"有完整表格吗？" → "吸烟点C1类有完整表格吗？"
+       - 上一轮问"三峡工程的投资情况"，追问"建设地点在哪？" → "三峡工程的建设地点在哪？"
+       - 上一轮问"货源投放有哪些原则？"，追问"具体内容是什么？" → "货源投放原则的具体内容是什么？"
+   - 如果问题本身已经完整且独立，直接返回原文
 
 2. **use_context**: 布尔值
    - true: 问题依赖历史对话中的信息，答案已经在历史回答中
@@ -105,7 +112,9 @@ class IntentAnalyzer:
    - 推理类（intent="reasoning"）：生成最多2个子查询
      * 原问题的检索查询
      * 一个补充角度的检索查询（如原因、背景、影响等），帮助获取更全面的上下文
-   - 其他类（factual/instruction/other）：严格只生成1个子查询（原问题）
+   - 其他类（factual/instruction/other）：严格只生成1个子查询
+     * 子查询应基于 rewritten_query（改写后的完整问题），而非用户原始输入
+     * 例如：追问"有完整表格吗？"改写为"吸烟点C1类有完整表格吗？"后，子查询应为"吸烟点C1类的完整表格内容"
    - 不要为同一实体生成语义重叠的查询
    - 子查询应保持原问题的关键词，长度20-60字符为宜
 
@@ -224,10 +233,10 @@ class IntentAnalyzer:
         self._exact_cache_max = 500
 
     def _get_client(self):
-        """获取 LLM 客户端"""
+        """获取 LLM 客户端（百炼快速模型）"""
         if self._client is None:
-            from config import get_llm_client
-            self._client = get_llm_client()
+            from config import get_intent_client
+            self._client = get_intent_client()
         return self._client
 
     def _get_cache(self):
@@ -299,17 +308,30 @@ class IntentAnalyzer:
             return self._exact_cache[exact_key]
 
         # 2. 尝试从语义缓存获取
+        #    关键：语义缓存只用原始 query 做 embedding（不含历史），
+        #    避免同会话中不同问题因历史上下文污染导致误命中
         cache = self._get_cache()
         if cache:
-            # 使用 query + 历史关键信息作为缓存键
-            cache_key = self._build_cache_key(query, history)
-            cache_emb = self._get_embedding(cache_key)
+            query_emb = self._get_embedding(query)
 
-            if cache_emb is not None:
-                cached = cache.get(cache_emb)
-                if cached:
+            if query_emb is not None:
+                cached = cache.get(query_emb)
+                # 确保缓存条目是意图分析结果（包含式校验，避免新增缓存类型时误命中）
+                if cached and cached.get("cache_type") == "intent_analysis":
+                    # 二次验证：检查原始 query 文本相似度
+                    cached_query = cached.get("_raw_query", "")
+                    if cached_query and self._query_text_similar(query, cached_query):
                         logger.info(f"意图分析缓存命中: {cached.get('reason', '')[:50]}")
                         return IntentAnalysis.from_dict(cached)
+                    elif cached_query:
+                        logger.info(
+                            f"意图分析缓存二次验证拒绝: "
+                            f"query='{query[:30]}' vs cached='{cached_query[:30]}'"
+                        )
+                    else:
+                        # 旧缓存无 _raw_query 字段，兼容放行
+                        logger.info(f"意图分析缓存命中(无验证): {cached.get('reason', '')[:50]}")
+                        return IntentAnalysis.from_dict(cached)
                 else:
                     logger.debug(f"意图分析缓存未命中，缓存状态: {cache.get_stats()}")
             else:
@@ -377,9 +399,13 @@ class IntentAnalyzer:
                     intent=intent_type
                 )
 
-                # 存入语义缓存
-                if cache and cache_emb is not None:
-                    cache.set(cache_emb, analysis.to_dict())
+                # 存入语义缓存（标记类型，避免与 RAG 回答缓存混淆）
+                # 使用仅含 query 的 embedding，不含历史，防止同会话误命中
+                if cache and query_emb is not None:
+                    cache_data = analysis.to_dict()
+                    cache_data["cache_type"] = "intent_analysis"
+                    cache_data["_raw_query"] = query  # 供二次验证使用
+                    cache.set(query_emb, cache_data)
 
                 # 存入精确匹配缓存
                 if len(self._exact_cache) < self._exact_cache_max:
@@ -419,6 +445,32 @@ class IntentAnalyzer:
 
         return " | ".join(parts)
 
+    @staticmethod
+    def _query_text_similar(query: str, cached_query: str, threshold: float = 0.5) -> bool:
+        """
+        判断两个 query 文本是否足够相似（字符级 Jaccard）。
+        用于语义缓存命中后的二次验证，防止语义相近但实际意图不同的问题误命中。
+
+        Args:
+            query: 当前查询
+            cached_query: 缓存中的原始查询
+            threshold: 相似度阈值，默认 0.5
+
+        Returns:
+            True 表示足够相似，可以命中缓存
+        """
+        # 精确匹配快速路径
+        if query.strip() == cached_query.strip():
+            return True
+        # 字符级 Jaccard 相似度
+        set_a = set(query)
+        set_b = set(cached_query)
+        if not set_a or not set_b:
+            return False
+        intersection = len(set_a & set_b)
+        union = len(set_a | set_b)
+        return (intersection / union) >= threshold
+
     def _build_history_summary(
         self,
         history: List[dict],
@@ -437,6 +489,7 @@ class IntentAnalyzer:
         if not history:
             return "（无历史对话）"
 
+        import re
         parts = []
 
         # 提取最近 3 轮对话
@@ -445,6 +498,7 @@ class IntentAnalyzer:
         for msg in recent_history:
             role = "用户" if msg.get("role") == "user" else "助手"
             content = msg.get("content", "")
+            original_content = content  # 保留原始内容用于结构化提取
 
             # 截断过长的内容
             if len(content) > 500:
@@ -452,9 +506,10 @@ class IntentAnalyzer:
 
             parts.append(f"【{role}】{content}")
 
-            # 提取图片信息
+            # 提取结构化信息（从 assistant 消息中提取章节、表格、来源等）
             metadata = msg.get("metadata", {})
             if isinstance(metadata, dict):
+                # 已有的图片提取
                 images = metadata.get("images", [])
                 if images:
                     for img in images[:3]:
@@ -463,6 +518,43 @@ class IntentAnalyzer:
                             img_type = img.get("type", "图片")
                             parts.append(f"  └─ {img_type}: {desc}")
 
+                # 来源文件提取
+                sources = metadata.get("sources", [])
+                if sources:
+                    source_names = []
+                    for s in sources[:3]:
+                        if isinstance(s, dict):
+                            name = s.get("source", "") or s.get("name", "")
+                            if name:
+                                source_names.append(name)
+                        elif isinstance(s, str):
+                            source_names.append(s)
+                    if source_names:
+                        parts.append(f"  └─ 来源文件: {', '.join(source_names)}")
+
+                # collections（检索知识库）提取
+                colls = metadata.get("collections", [])
+                if colls:
+                    parts.append(f"  └─ 检索知识库: {', '.join(colls)}")
+
+            # 从 assistant 原始内容中提取章节路径和表格结构
+            if role == "助手" and original_content:
+                # 提取章节路径：━ xxx ━ 格式
+                sections = re.findall(r'━\s*(.+?)\s*━', original_content)
+                if sections:
+                    unique_sections = list(dict.fromkeys(sections))  # 去重保序
+                    parts.append(f"  └─ 涉及章节: {'; '.join(unique_sections[:3])}")
+
+                # 提取表格列名：| A | B | C | 格式的表头行
+                table_headers = re.findall(r'^\|\s*(.+?)\s*\|', original_content, re.MULTILINE)
+                if table_headers:
+                    # 取第一个表格的列名
+                    first_header = table_headers[0]
+                    cols = [c.strip() for c in first_header.split('|') if c.strip()]
+                    # 排除分隔符行（--- 格式）
+                    if cols and not all(re.match(r'^[-:]+$', c) for c in cols):
+                        parts.append(f"  └─ 含表格，列名: {', '.join(cols[:6])}")
+
         # 添加图片上下文
         if context_images:
             parts.append("\n【上下文中的图片】")

core/llm_utils.py

@@ -72,16 +72,35 @@ def call_llm(
 
         content = response.choices[0].message.content
         
-        # 推理模型兼容：content 为空时尝试从 reasoning_content 提取
+        # 推理模型兼容（mimo-v2.5 等）：
+        # 推理模型思考链消耗大量 token（~1000），max_tokens 不足时 content 为空，
+        # 全部输出进入 reasoning_content。此处从思考链中提取有效内容。
         if not content or not content.strip():
             reasoning = getattr(response.choices[0].message, 'reasoning_content', None)
             if reasoning and reasoning.strip():
-                # 从思维链中提取 JSON 块作为内容
-                json_match = re.search(r'\{[\s\S]*\}', reasoning)
+                # 先去掉 <think>...</think> 标签
+                cleaned = re.sub(r'', '', reasoning, flags=re.DOTALL).strip()
+                if cleaned:
+                    logger.info("LLM: content为空，从reasoning_content提取内容")
+                    # 尝试提取 JSON 对象（兼容结构化响应场景）
+                    json_match = re.search(r'\{[\s\S]*\}', cleaned)
                     if json_match:
-                    logger.info("LLM: content为空，从reasoning_content提取JSON")
+                        try:
+                            json.loads(json_match.group())
                             return json_match.group().strip()
-            logger.warning("LLM 返回空 content（可能需要增大 max_tokens）")
+                        except (json.JSONDecodeError, ValueError):
+                            pass
+                    # 尝试提取 JSON 数组
+                    bracket_match = re.search(r'\[[\s\S]*\]', cleaned)
+                    if bracket_match:
+                        try:
+                            json.loads(bracket_match.group())
+                            return bracket_match.group().strip()
+                        except (json.JSONDecodeError, ValueError):
+                            pass
+                    # 纯文本响应：直接返回清理后的内容
+                    return cleaned
+            logger.warning("LLM 返回空 content 且 reasoning_content 也无法提取（可能需要增大 max_tokens）")
             return None
         
         return content.strip()
@@ -95,7 +114,7 @@ def call_llm_stream(
     prompt: str,
     model: str,
     temperature: float = 0.3,
-    max_tokens: int = 1000,
+    max_tokens: int = 3000,
     messages: List[dict] = None,
     error_prefix: str = "[错误]",
     **kwargs
@@ -104,13 +123,14 @@ def call_llm_stream(
     流式 LLM 调用（生成器封装）
 
     自动处理流式响应，逐块 yield 文本内容。
+    兼容推理模型（mimo-v2.5 等）：当 content 为空时回退到 reasoning_content。
 
     Args:
         client: OpenAI 客户端实例
         prompt: 用户提示
         model: 模型名称
         temperature: 温度参数
-        max_tokens: 最大 token 数
+        max_tokens: 最大 token 数（推理模型需留足思考链预算）
         messages: 完整消息列表
         error_prefix: 错误时的前缀
         **kwargs: 其他参数
@@ -135,9 +155,33 @@ def call_llm_stream(
             **kwargs
         )
 
+        content_yielded = False
+        reasoning_buffer = []
+
         for chunk in stream:
-            if chunk.choices and chunk.choices[0].delta.content:
-                yield chunk.choices[0].delta.content
+            if not chunk.choices:
+                continue
+            delta = chunk.choices[0].delta
+
+            # 正常 content 输出
+            if hasattr(delta, 'content') and delta.content:
+                content_yielded = True
+                yield delta.content
+                continue
+
+            # 推理模型：reasoning_content（思考链）
+            rc = getattr(delta, 'reasoning_content', None)
+            if rc:
+                reasoning_buffer.append(rc)
+
+        # 回退：content 为空但 reasoning_content 有内容（推理模型 token 不足时）
+        if not content_yielded and reasoning_buffer:
+            reasoning_text = ''.join(reasoning_buffer)
+            # 去掉 <think>...</think> 标签
+            cleaned = re.sub(r'', '', reasoning_text, flags=re.DOTALL).strip()
+            if cleaned:
+                logger.info("流式 LLM: content为空，从reasoning_content提取内容")
+                yield cleaned
 
     except Exception as e:
         logger.error(f"LLM 流式调用失败: {e}")
@@ -352,7 +396,7 @@ def quick_yes_no(
     if keywords is None:
         keywords = ["是", "需要", "yes", "true"]
 
-    result = call_llm(client, prompt, model, temperature=0, max_tokens=10)
+    result = call_llm(client, prompt, model, temperature=0, max_tokens=128)
     if result is None:
         return False
 

core/mmr.py

@@ -108,22 +108,44 @@ def mmr_rerank(
     return selected
 
 
+def _tokenize_words(text: str) -> set:
+    """
+    使用 jieba 分词并过滤噪声，返回有意义的词集合。
+
+    过滤规则：
+    - 去除单字符词（如 "的", "了", "在"）—— 这些是停用词，对区分文档无意义
+    - 去除纯数字 / 纯标点
+    - 保留 2 字及以上的实词
+    """
+    import jieba
+    words = set()
+    for w in jieba.cut(text):
+        w = w.strip()
+        if len(w) >= 2 and not w.isdigit():
+            words.add(w)
+    return words
+
+
 def mmr_filter_by_content(
     candidates: List[Dict],
     top_k: int = 30,
-    similarity_threshold: float = 0.9
+    similarity_threshold: float = 0.85
 ) -> List[Dict]:
     """
-    基于内容相似度的去重（简化版，不需要 embedding）
+    基于 jieba 词级 Jaccard 相似度的去重（不需要 embedding）
+
+    与旧版字符级 set(text) 的区别：
+    - 旧版：set("安全生产管理制度") → {'安','全','生','产',...}，中文文档间字符集合高度重叠
+    - 新版：jieba 分词 → {"安全生产", "管理制度", ...}，词级集合区分度高
 
     适用于：
-    - 没有 embedding 的情况
-    - 快速去重场景
+    - MMR_USE_EMBEDDING=False 时的快速去重
+    - 避免 CPU 编码 100+ 文档的 50 秒开销
 
     Args:
         candidates: 候选文档列表
         top_k: 返回数量
-        similarity_threshold: 相似度阈值，超过则视为重复
+        similarity_threshold: 相似度阈值，超过则视为重复（默认 0.85）
 
     Returns:
         去重后的候选文档列表
@@ -134,25 +156,32 @@ def mmr_filter_by_content(
     if len(candidates) <= top_k:
         return candidates
 
-    selected = []
-    remaining = candidates.copy()
+    # 预分词：对所有候选文档一次性分词，避免重复调用 jieba.cut
+    word_sets = []
+    for c in candidates:
+        content = c.get('content', c.get('document', ''))[:500]
+        word_sets.append(_tokenize_words(content))
 
-    while len(selected) < top_k and remaining:
-        current = remaining.pop(0)
+    selected_indices = []
+
+    for i in range(len(candidates)):
+        if len(selected_indices) >= top_k:
+            break
+
+        current_words = word_sets[i]
+        if not current_words:
+            # 空内容直接保留
+            selected_indices.append(i)
+            continue
 
-        # 检查是否与已选内容重复
         is_duplicate = False
-        current_content = current.get('content', current.get('document', ''))[:200]
+        for j in selected_indices:
+            selected_words = word_sets[j]
+            if not selected_words:
+                continue
 
-        for s in selected:
-            s_content = s.get('content', s.get('document', ''))[:200]
-
-            # 简单的 Jaccard 相似度
-            words1 = set(current_content)
-            words2 = set(s_content)
-            if words1 and words2:
-                intersection = len(words1 & words2)
-                union = len(words1 | words2)
+            intersection = len(current_words & selected_words)
+            union = len(current_words | selected_words)
             similarity = intersection / union if union > 0 else 0
 
             if similarity > similarity_threshold:
@@ -160,9 +189,9 @@ def mmr_filter_by_content(
                 break
 
         if not is_duplicate:
-            selected.append(current)
+            selected_indices.append(i)
 
-    return selected
+    return [candidates[i] for i in selected_indices]
 
 
 # ==================== 测试 ====================

knowledge/base.py

@@ -382,18 +382,34 @@ class BM25Index:
 
     def add_documents(self, ids: List[str], documents: List[str], metadatas: List[dict]) -> None:
         """
-        添加文档到索引（会覆盖原有索引）
+        添加文档到索引（追加模式，自动去重）
+
+        如果 ID 已存在则更新对应文档，否则追加新文档。
+        添加后自动重建 BM25 索引。
 
         Args:
             ids: 文档 ID 列表
             documents: 文档内容列表
             metadatas: 文档元数据列表
         """
-        self.ids = ids
-        self.documents = documents
-        self.metadatas = metadatas
-        if documents:
-            tokenized = [self.tokenize(doc) for doc in documents]
+        # 建立已有 ID -> 索引位置 的映射，用于去重
+        existing_map = {doc_id: idx for idx, doc_id in enumerate(self.ids)}
+
+        for i, doc_id in enumerate(ids):
+            if doc_id in existing_map:
+                # 更新已有文档
+                pos = existing_map[doc_id]
+                self.documents[pos] = documents[i]
+                self.metadatas[pos] = metadatas[i]
+            else:
+                # 追加新文档
+                existing_map[doc_id] = len(self.ids)
+                self.ids.append(doc_id)
+                self.documents.append(documents[i])
+                self.metadatas.append(metadatas[i])
+
+        if self.documents:
+            tokenized = [self.tokenize(doc) for doc in self.documents]
             self.bm25 = BM25Okapi(tokenized)
 
     def search(self, query: str, top_k: int = 10) -> Tuple[List[str], List[str], List[dict], List[float]]:

knowledge/collection.py

@@ -134,6 +134,9 @@ class CollectionMixin:
         """
         from .base import BM25Index
 
+        # 从磁盘重新加载元数据，确保多 worker 进程间状态一致
+        self._metadata = self._load_metadata()
+
         if not kb_name or not kb_name.replace('_', '').isalnum():
             return False, "向量库名称只能包含字母、数字和下划线"
 
@@ -190,6 +193,9 @@ class CollectionMixin:
         Returns:
             更新成功返回 True，向量库不存在返回 False
         """
+        # 从磁盘重新加载元数据，确保多 worker 进程间状态一致
+        self._metadata = self._load_metadata()
+
         collections = self._metadata.get("collections", {})
         if kb_name not in collections:
             return False
@@ -228,6 +234,9 @@ class CollectionMixin:
         """
         import shutil
 
+        # 从磁盘重新加载元数据，确保多 worker 进程间状态一致
+        self._metadata = self._load_metadata()
+
         if kb_name == PUBLIC_KB_NAME:
             return False, "公开知识库不能删除"
 
@@ -322,6 +331,21 @@ class CollectionMixin:
             except Exception as e:
                 logger.warning(f"清理版本记录失败: {e}")
 
+            # 清理不再被引用的图片和 VLM 缓存文件
+            # 注意：此时 ChromaDB collection 已删除，cleanup_image_orphans 会扫描
+            # 所有剩余 collection，仅该 collection 引用的图片会被识别为孤儿
+            try:
+                from knowledge.image_cleanup import cleanup_image_orphans
+                cleanup_result = cleanup_image_orphans(self)
+                if cleanup_result['deleted_images'] or cleanup_result['deleted_caches']:
+                    logger.info(
+                        f"清理孤儿文件: {cleanup_result['deleted_images']} 图片 + "
+                        f"{cleanup_result['deleted_caches']} VLM缓存, "
+                        f"释放 {cleanup_result['freed_bytes']/1024:.1f} KB"
+                    )
+            except Exception as e:
+                logger.warning(f"清理孤儿文件失败: {e}")
+
             if kb_name in self._metadata.get("collections", {}):
                 del self._metadata["collections"][kb_name]
             self._save_metadata()
@@ -348,6 +372,8 @@ class CollectionMixin:
             - department: 所属部门
             - description: 描述
         """
+        # 从磁盘重新加载元数据，确保多 worker 进程间状态一致
+        self._metadata = self._load_metadata()
         result = []
 
         # 扫描 base_path 下的所有子目录作为向量库
@@ -382,7 +408,10 @@ class CollectionMixin:
 
         self._save_metadata()
 
+        stale_collections = []
+
         for name, info in self._metadata.get("collections", {}).items():
+            try:
                 collection = self.get_collection(name)
                 result.append(CollectionInfo(
                     name=name,
@@ -392,6 +421,18 @@ class CollectionMixin:
                     department=info.get("department", ""),
                     description=info.get("description", "")
                 ))
+            except Exception as e:
+                logger.warning(
+                    f"跳过异常向量库 '{name}': {e}，可能是 ChromaDB 数据目录已丢失"
+                )
+                stale_collections.append(name)
+
+        # 清理元数据中指向已失效集合的条目
+        if stale_collections:
+            for name in stale_collections:
+                self._metadata.get("collections", {}).pop(name, None)
+                logger.info(f"清理失效向量库元数据: {name}")
+            self._save_metadata()
 
         return result
 
@@ -405,4 +446,6 @@ class CollectionMixin:
         Returns:
             存在返回 True，不存在返回 False
         """
+        # 从磁盘重新加载元数据，确保多 worker 进程间状态一致
+        self._metadata = self._load_metadata()
         return kb_name in self._metadata.get("collections", {})

knowledge/document.py

@@ -72,6 +72,18 @@ class DocumentMixin:
         except Exception as e:
             logger.warning(f"清理版本记录失败: {e}")
 
+        # 清理不再被引用的图片和 VLM 缓存文件
+        try:
+            from knowledge.image_cleanup import cleanup_image_orphans
+            cleanup_result = cleanup_image_orphans(self, collections=[kb_name])
+            if cleanup_result['deleted_images'] or cleanup_result['deleted_caches']:
+                logger.info(
+                    f"清理孤儿文件: {cleanup_result['deleted_images']} 图片 + "
+                    f"{cleanup_result['deleted_caches']} VLM缓存"
+                )
+        except Exception as e:
+            logger.warning(f"清理孤儿文件失败: {e}")
+
         logger.info(f"从 {kb_name} 删除文档: {filename}, 片段数: {deleted}")
         return deleted
 

knowledge/image_cleanup.py

@@ -0,0 +1,140 @@
+"""
+图片/VLM缓存孤儿文件清理模块
+
+提供可被 document.py / collection.py 调用的清理函数，
+也可被 cleanup_orphans.py 独立脚本使用。
+"""
+
+import hashlib
+import logging
+import os
+from pathlib import Path
+
+logger = logging.getLogger(__name__)
+
+IMAGES_DIR = Path(".data/images")
+VLM_CACHE_DIR = Path(".data/cache/vlm")
+
+
+def compute_file_hash(file_path: str) -> str:
+    """计算文件 MD5"""
+    with open(file_path, 'rb') as f:
+        return hashlib.md5(f.read()).hexdigest()
+
+
+def collect_referenced_images(manager, collections=None) -> set:
+    """
+    从 ChromaDB 收集所有被引用的图片文件名。
+
+    Args:
+        manager: KnowledgeBaseManager 实例
+        collections: 限定知识库列表，None 表示全部
+
+    Returns:
+        set of image filenames (e.g., {"185a7a75d246.png", ...})
+    """
+    referenced = set()
+
+    if collections:
+        kb_names = collections
+    else:
+        kb_names = [c.name if hasattr(c, 'name') else str(c)
+                     for c in manager.list_collections()]
+
+    for kb_name in kb_names:
+        try:
+            col = manager.get_collection(kb_name)
+        except Exception:
+            continue
+
+        results = col.get(include=['metadatas'])
+        if not results['ids']:
+            continue
+
+        for meta in results['metadatas']:
+            image_path = meta.get('image_path', '')
+            if image_path:
+                referenced.add(os.path.basename(image_path))
+
+    return referenced
+
+
+def cleanup_image_orphans(manager, collections=None, dry_run=False) -> dict:
+    """
+    清理不再被任何 ChromaDB 切片引用的图片和 VLM 缓存文件。
+
+    Args:
+        manager: KnowledgeBaseManager 实例
+        collections: 限定知识库列表，None 表示全部
+        dry_run: True 时只返回孤儿列表不实际删除
+
+    Returns:
+        {
+            'orphan_images': [(filepath, filename, size_bytes)],
+            'orphan_caches': [(filepath, filename, size_bytes)],
+            'deleted_images': int,
+            'deleted_caches': int,
+            'freed_bytes': int
+        }
+    """
+    result = {
+        'orphan_images': [],
+        'orphan_caches': [],
+        'deleted_images': 0,
+        'deleted_caches': 0,
+        'freed_bytes': 0
+    }
+
+    # 1. 收集引用
+    referenced = collect_referenced_images(manager, collections)
+
+    # 2. 查找孤儿图片
+    if IMAGES_DIR.exists():
+        for f in IMAGES_DIR.iterdir():
+            if f.is_file() and f.name not in referenced:
+                result['orphan_images'].append((str(f), f.name, f.stat().st_size))
+
+    # 3. 查找孤儿 VLM 缓存（图片已删除则缓存也应是孤儿）
+    if VLM_CACHE_DIR.exists():
+        referenced_hashes = set()
+        for filename in referenced:
+            full_path = IMAGES_DIR / filename
+            if full_path.exists():
+                try:
+                    img_hash = compute_file_hash(str(full_path))
+                    referenced_hashes.add(img_hash)
+                except Exception:
+                    pass
+
+        for f in VLM_CACHE_DIR.iterdir():
+            if f.is_file() and f.suffix == '.txt':
+                cache_hash = f.stem
+                if cache_hash not in referenced_hashes:
+                    result['orphan_caches'].append((str(f), f.name, f.stat().st_size))
+
+    # 4. 删除
+    if not dry_run:
+        for filepath, filename, size in result['orphan_images']:
+            try:
+                os.remove(filepath)
+                result['deleted_images'] += 1
+                result['freed_bytes'] += size
+            except OSError as e:
+                logger.warning(f"删除孤儿图片失败: {filename} - {e}")
+
+        for filepath, filename, size in result['orphan_caches']:
+            try:
+                os.remove(filepath)
+                result['deleted_caches'] += 1
+                result['freed_bytes'] += size
+            except OSError as e:
+                logger.warning(f"删除孤儿缓存失败: {filename} - {e}")
+
+        if result['deleted_images'] or result['deleted_caches']:
+            logger.info(
+                f"清理孤儿: {result['deleted_images']} 图片 + "
+                f"{result['deleted_caches']} 缓存, "
+                f"释放 {result['freed_bytes']/1024:.1f} KB"
+            )
+
+    return result

knowledge/lazy_enhance.py

@@ -25,7 +25,20 @@ def compute_file_hash(file_path: str) -> str:
         return hashlib.md5(file_path.encode()).hexdigest()
 
 
-async def lazy_vlm_description(chunk_id: str, image_path: str, kb_name: str, metadata: dict = None) -> str:
+def _get_embedding_model():
+    """从 RAGEngine 获取 embedding 模型（KnowledgeBaseManager 上没有此属性）"""
+    try:
+        from core.engine import get_engine
+        engine = get_engine()
+        if not engine._initialized:
+            engine.initialize()
+        return engine.embedding_model
+    except Exception as e:
+        logger.warning(f"获取 embedding 模型失败: {e}")
+        return None
+
+
+async def lazy_vlm_description(chunk_id: str, image_path: str, kb_name: str, metadata: dict = None, defer_chromadb: bool = False) -> str:
     """
     懒加载 VLM 描述
 
@@ -36,6 +49,7 @@ async def lazy_vlm_description(chunk_id: str, image_path: str, kb_name: str, met
         image_path: 图片路径（相对路径或绝对路径）
         kb_name: 知识库名称
         metadata: 图片元数据（包含 section、page、caption、上下文等）
+        defer_chromadb: 为 True 时跳过 ChromaDB 更新（仅写文件缓存），避免后台线程写锁竞争
 
     Returns:
         VLM 生成的图片描述
@@ -49,23 +63,45 @@ async def lazy_vlm_description(chunk_id: str, image_path: str, kb_name: str, met
     else:
         full_image_path = image_path
 
-    # 1. 检查缓存
+    # 1. 检查缓存（空缓存视为无效，需重新生成）
     img_hash = compute_file_hash(full_image_path)
     cache_file = VLM_CACHE_DIR / f"{img_hash}.txt"
     if cache_file.exists():
+        cached = cache_file.read_text(encoding='utf-8')
+        if len(cached.strip()) >= 5:
             logger.info(f"VLM 缓存命中: {image_path}")
-        return cache_file.read_text(encoding='utf-8')
+            return cached
+        else:
+            logger.warning(f"VLM 缓存内容过短({len(cached.strip())}字符)，删除并重新生成: {image_path}")
+            try:
+                cache_file.unlink()
+            except OSError:
+                pass
 
     # 2. 调用 VLM（传入元数据）
     logger.info(f"VLM 懒加载: {image_path}")
     kb_manager = get_kb_manager()
     description = kb_manager._generate_image_description(full_image_path, metadata=metadata)
 
-    # 3. 写入缓存
+    # 3. 空描述保护：VLM 返回内容过短时不写入缓存和向量库
+    if not description or len(description.strip()) < 5:
+        logger.warning(f"VLM 返回描述过短({len(description.strip()) if description else 0}字符)，跳过缓存和向量库更新: {image_path}")
+        return description or ''
+
+    # 4. 写入缓存
     VLM_CACHE_DIR.mkdir(parents=True, exist_ok=True)
     cache_file.write_text(description, encoding='utf-8')
 
-    # 4. 更新向量库（metadata + embedding）
+    # 5. 更新向量库（metadata + embedding），需校验 chunk_id 非空
+    # defer_chromadb=True 时跳过（后台线程只写缓存，避免 SQLite 写锁竞争）
+    if defer_chromadb:
+        logger.info(f"延迟 ChromaDB 更新(仅写缓存): {chunk_id}")
+        return description
+
+    if not chunk_id:
+        logger.warning("chunk_id 为空，跳过向量库更新")
+        return description
+
     try:
         collection = kb_manager.get_collection(kb_name)
         result = collection.get(ids=[chunk_id], include=['metadatas'])
@@ -79,7 +115,7 @@ async def lazy_vlm_description(chunk_id: str, image_path: str, kb_name: str, met
 
             # 更新 embedding（使用 VLM 描述重新计算向量）
             # 这样 VLM 描述中的关键词（如"发电量"）才能参与相似度检索
-            embedding_model = kb_manager.embedding_model
+            embedding_model = _get_embedding_model()
             if embedding_model:
                 new_vector = embedding_model.encode(description).tolist()
                 if isinstance(new_vector[0], list):
@@ -91,20 +127,21 @@ async def lazy_vlm_description(chunk_id: str, image_path: str, kb_name: str, met
                     embeddings=[new_vector],
                     documents=[description]  # 同时更新 document 字段
                 )
-                logger.info(f"已更新向量库 embedding: {chunk_id}")
+                logger.info(f"已更新向量库(embedding+metadata): {chunk_id}")
             else:
                 # 无 embedding 模型时只更新 metadata
                 collection.update(
                     ids=[chunk_id],
                     metadatas=[new_metadata]
                 )
+                logger.info(f"已更新向量库(仅metadata,无embedding模型): {chunk_id}")
     except Exception as e:
         logger.warning(f"更新向量库失败: {e}")
 
     return description
 
 
-async def lazy_table_summary(chunk_id: str, table_md: str, kb_name: str) -> str:
+async def lazy_table_summary(chunk_id: str, table_md: str, kb_name: str, defer_chromadb: bool = False) -> str:
     """
     懒加载表格摘要
 
@@ -114,35 +151,58 @@ async def lazy_table_summary(chunk_id: str, table_md: str, kb_name: str) -> str:
         chunk_id: 切片 ID
         table_md: 表格 Markdown 内容
         kb_name: 知识库名称
+        defer_chromadb: 为 True 时跳过 ChromaDB 更新（仅写文件缓存），避免后台线程写锁竞争
 
     Returns:
         LLM 生成的表格摘要
     """
     from knowledge.manager import get_kb_manager
 
-    # 1. 检查缓存
+    # 1. 检查缓存（空缓存视为无效）
     table_hash = hashlib.md5(table_md.encode()).hexdigest()
     cache_file = LLM_CACHE_DIR / f"{table_hash}.txt"
     if cache_file.exists():
+        cached = cache_file.read_text(encoding='utf-8')
+        if len(cached.strip()) >= 5:
             logger.info(f"LLM 缓存命中: {chunk_id}")
-        return cache_file.read_text(encoding='utf-8')
+            return cached
+        else:
+            logger.warning(f"LLM 缓存内容过短({len(cached.strip())}字符)，删除并重新生成: {chunk_id}")
+            try:
+                cache_file.unlink()
+            except OSError:
+                pass
 
     # 2. 调用 LLM
     logger.info(f"LLM 懒加载: {chunk_id}")
     kb_manager = get_kb_manager()
     summary = kb_manager._generate_table_summary(table_md, None)
 
+    # 空摘要保护
+    if not summary or len(summary.strip()) < 5:
+        logger.warning(f"LLM 返回摘要过短，跳过缓存和向量库更新: {chunk_id}")
+        return summary or ''
+
     # 3. 写入缓存
     LLM_CACHE_DIR.mkdir(parents=True, exist_ok=True)
     cache_file.write_text(summary, encoding='utf-8')
 
-    # 4. 更新向量库（可选）
+    # 4. 更新向量库，需校验 chunk_id 非空
+    # defer_chromadb=True 时跳过（后台线程只写缓存，避免 SQLite 写锁竞争）
+    if defer_chromadb:
+        logger.info(f"延迟 ChromaDB 更新(仅写缓存): {chunk_id}")
+        return summary
+
+    if not chunk_id:
+        logger.warning("chunk_id 为空，跳过表格向量库更新")
+        return summary
     try:
         collection = kb_manager.get_collection(kb_name)
         result = collection.get(ids=[chunk_id], include=['metadatas'])
         if result['metadatas']:
-            # 新增摘要切片
-            embedding_model = kb_manager.embedding_model
+            # 新增摘要切片（需要 embedding 模型）
+            embedding_model = _get_embedding_model()
+            if embedding_model:
                 vector = embedding_model.encode(summary).tolist()
                 if isinstance(vector[0], list):
                     vector = vector[0]
@@ -157,7 +217,10 @@ async def lazy_table_summary(chunk_id: str, table_md: str, kb_name: str) -> str:
                         'original_doc_id': chunk_id
                     }]
                 )
-            # 更新原切片标记
+                logger.info(f"已新增摘要切片(embedding): {chunk_id}_summary")
+            else:
+                logger.info(f"跳过摘要切片(无embedding模型): {chunk_id}")
+            # 更新原切片标记（不依赖 embedding 模型）
             collection.update(
                 ids=[chunk_id],
                 metadatas=[{**result['metadatas'][0], 'has_summary': True}]
@@ -168,7 +231,7 @@ async def lazy_table_summary(chunk_id: str, table_md: str, kb_name: str) -> str:
     return summary
 
 
-async def enhance_retrieved_chunks(contexts: list, query: str, kb_name: str):
+async def enhance_retrieved_chunks(contexts: list, query: str, kb_name: str, defer_chromadb: bool = False):
     """
     检索后增强：按需调用 LLM/VLM
 
@@ -176,8 +239,12 @@ async def enhance_retrieved_chunks(contexts: list, query: str, kb_name: str):
         contexts: 检索上下文列表
         query: 用户查询
         kb_name: 知识库名称
+        defer_chromadb: 为 True 时后台线程只写文件缓存，不更新 ChromaDB（避免写锁竞争）
     """
+    import re
+
     for ctx in contexts:
+        try:
             meta = ctx.get('meta', {})
             chunk_type = meta.get('chunk_type', 'text')
             image_path = meta.get('image_path', '')
@@ -185,14 +252,11 @@ async def enhance_retrieved_chunks(contexts: list, query: str, kb_name: str):
             # 图片切片：懒加载 VLM 描述
             if chunk_type in ('image', 'chart') and not meta.get('has_vlm_desc'):
                 if image_path:
-                try:
                     # 从 doc 字段中提取图号（上下文可能包含"见图2.5"等）
                     doc_text = ctx.get('doc', '')
-                    import re
 
                     # 提取图号（从前文/后文中）
                     figure_number = ""
-                    # 匹配 "见图2.5"、"图2.5"、"见图 2.5" 等
                     fig_match = re.search(r'[见如]?图\s*(\d+\.?\d*)', doc_text)
                     if fig_match:
                         figure_number = fig_match.group(1)
@@ -210,19 +274,19 @@ async def enhance_retrieved_chunks(contexts: list, query: str, kb_name: str):
                         'page': meta.get('page'),
                         'caption': meta.get('caption', ''),
                         'source': meta.get('source', ''),
-                        'figure_number': figure_number,  # 添加提取的图号
-                        'doc_text': doc_text  # 添加完整文档文本
+                        'figure_number': figure_number,
+                        'doc_text': doc_text
                     }
                     vlm_desc = await lazy_vlm_description(
-                        meta.get('id', ''),
+                        meta.get('chunk_id', ''),
                         image_path,
                         kb_name,
-                        metadata=image_metadata
+                        metadata=image_metadata,
+                        defer_chromadb=defer_chromadb
                     )
+                    if vlm_desc:
                         ctx['doc'] = vlm_desc
                         ctx['vlm_enhanced'] = True
-                except Exception as e:
-                    logger.warning(f"VLM 懒加载失败: {e}")
 
             # 表格切片：同时处理摘要和关联图片的 VLM 描述
             elif chunk_type == 'table':
@@ -230,58 +294,53 @@ async def enhance_retrieved_chunks(contexts: list, query: str, kb_name: str):
 
                 # 1. 懒加载表格摘要（高分切片）
                 if not meta.get('has_summary'):
-                score = meta.get('score', 0)
-                if score > 0.7:  # 只对高相关表格生成摘要
-                    try:
+                    score = ctx.get('score', 0)
+                    if score > 0.7:
                         summary = await lazy_table_summary(
-                            meta.get('id', ''),
+                            meta.get('chunk_id', ''),
                             doc_text,
-                            kb_name
+                            kb_name,
+                            defer_chromadb=defer_chromadb
                         )
-                        # 摘要作为补充信息
+                        if summary:
                             ctx['summary'] = summary
                             ctx['llm_enhanced'] = True
-                    except Exception as e:
-                        logger.warning(f"表格摘要懒加载失败: {e}")
 
                 # 2. 表格有关联图片时，懒加载 VLM 描述
                 if image_path and not meta.get('has_vlm_desc'):
-                try:
-                    import re
-
                     # 提取表号（如 "表2.2"、"见表2.1"）
                     table_number = ""
-                    # 匹配 "表2.2"、"见表2.2"、"见表 2.2" 等
                     table_match = re.search(r'[见如]?表\s*(\d+\.?\d*)', doc_text)
                     if table_match:
                         table_number = table_match.group(1)
 
-                    # 如果 doc 中没有，尝试从 section 中提取
                     section = meta.get('section') or meta.get('section_path', '')
                     if not table_number and section:
                         table_match = re.search(r'[见如]?表\s*(\d+\.?\d*)', section)
                         if table_match:
                             table_number = table_match.group(1)
 
-                    # 构建表格图片元数据
                     table_image_metadata = {
                         'section': section,
                         'page': meta.get('page'),
                         'caption': meta.get('caption', ''),
                         'source': meta.get('source', ''),
-                        'table_number': table_number,  # 表号
-                        'figure_number': table_number,  # 兼容字段
+                        'table_number': table_number,
+                        'figure_number': table_number,
                         'doc_text': doc_text,
-                        'is_table': True  # 标记为表格图片
+                        'is_table': True
                     }
                     vlm_desc = await lazy_vlm_description(
-                        meta.get('id', ''),
+                        meta.get('chunk_id', ''),
                         image_path,
                         kb_name,
-                        metadata=table_image_metadata
+                        metadata=table_image_metadata,
+                        defer_chromadb=defer_chromadb
                     )
-                    # 表格图片描述作为补充信息
+                    if vlm_desc:
                         ctx['image_description'] = vlm_desc
                         ctx['vlm_enhanced'] = True
+
         except Exception as e:
-                    logger.warning(f"表格图片 VLM 懒加载失败: {e}")
+            chunk_id = ctx.get('meta', {}).get('chunk_id', '?')
+            logger.warning(f"增强切片失败(chunk_id={chunk_id}): {e}")

knowledge/manager.py

@@ -29,11 +29,17 @@
 import os
 import json
 import threading
+try:
+    import fcntl
+    _HAS_FCNTL = True
+except ImportError:
+    _HAS_FCNTL = False  # Windows 环境无 fcntl
 from typing import List, Dict, Optional, Tuple
 from pathlib import Path
 import logging
 
 import chromadb
+from bs4 import BeautifulSoup
 
 # 从 base.py 导入基础类和常量
 from .base import (
@@ -134,22 +140,36 @@ class KnowledgeBaseManager(
         logger.info(f"知识库管理器初始化完成，路径: {self.base_path}，发现 {len(existing_kbs)} 个向量库: {existing_kbs}")
 
     def _load_metadata(self) -> dict:
-        """加载元数据"""
+        """加载元数据（带文件锁，确保多 worker 进程间一致）"""
         metadata_path = os.path.join(self.base_path, KB_METADATA_FILE)
         if os.path.exists(metadata_path):
             try:
                 with open(metadata_path, 'r', encoding='utf-8') as f:
+                    if _HAS_FCNTL:
+                        fcntl.flock(f.fileno(), fcntl.LOCK_SH)
+                    try:
                         return json.load(f)
+                    finally:
+                        if _HAS_FCNTL:
+                            fcntl.flock(f.fileno(), fcntl.LOCK_UN)
             except Exception as e:
                 logger.error(f"加载元数据失败: {e}")
         return {"collections": {}}
 
     def _save_metadata(self):
-        """保存元数据"""
+        """保存元数据（带文件锁，防止并发写入数据覆盖）"""
         metadata_path = os.path.join(self.base_path, KB_METADATA_FILE)
         try:
             with open(metadata_path, 'w', encoding='utf-8') as f:
+                if _HAS_FCNTL:
+                    fcntl.flock(f.fileno(), fcntl.LOCK_EX)
+                try:
                     json.dump(self._metadata, f, ensure_ascii=False, indent=2)
+                    f.flush()
+                    os.fsync(f.fileno())
+                finally:
+                    if _HAS_FCNTL:
+                        fcntl.flock(f.fileno(), fcntl.LOCK_UN)
         except Exception as e:
             logger.error(f"保存元数据失败: {e}")
 
@@ -316,6 +336,7 @@ class KnowledgeBaseManager(
                 "section": section_path,
                 "status": "active",
                 "version": "v1",
+                "text_level": getattr(chunk, 'text_level', 0),  # 标题级别(0=正文,1=h1,2=h2,3=h3)，供检索管线层次感知
             }
 
             if extra_metadata:
@@ -328,6 +349,25 @@ class KnowledgeBaseManager(
             if hasattr(chunk, 'image_path') and chunk.image_path:
                 metadata['image_path'] = chunk.image_path
 
+            # bbox 坐标（PDF 有，DOCX 无，需 None 保护）
+            # _build_citation() 从 metadata 读取 bbox 做引用定位
+            chunk_bbox = getattr(chunk, 'bbox', None)
+            if chunk_bbox:
+                metadata['bbox'] = json.dumps(chunk_bbox)
+
+            # MinerU 结构化元数据（表格类型、嵌套层级、图片子类型）
+            chunk_table_type = getattr(chunk, 'table_type', '')
+            if chunk_table_type:
+                metadata['table_type'] = chunk_table_type
+
+            chunk_nest_level = getattr(chunk, 'table_nest_level', '')
+            if chunk_nest_level:
+                metadata['table_nest_level'] = str(chunk_nest_level)
+
+            chunk_sub_type = getattr(chunk, 'sub_type', '')
+            if chunk_sub_type:
+                metadata['sub_type'] = chunk_sub_type
+
             # 生成向量
             try:
                 embedding = embedding_model.encode(semantic_content).tolist()
@@ -393,6 +433,11 @@ class KnowledgeBaseManager(
         if len(chunks) < 2:
             return chunks
 
+        # 检测页码是否可靠：若所有 chunk 的 page_start 相同（如 Word 文档 page_idx 全为 0），
+        # 则页码信息不可用，需要启用降级合并规则
+        page_values = set(getattr(c, 'page_start', 0) for c in chunks)
+        pages_unavailable = len(page_values) <= 1
+
         merged_chunks = []
         i = 0
         merge_count = 0
@@ -405,6 +450,7 @@ class KnowledgeBaseManager(
                 # 查找下一个表格（跳过中间的"续表"文本）
                 next_table_idx = None
                 next_chunk = None
+                intermediate_texts = []  # 收集中间文本用于降级判断
 
                 for j in range(i + 1, min(i + 4, len(chunks))):  # 最多向前看3个切片
                     candidate = chunks[j]
@@ -419,7 +465,11 @@ class KnowledgeBaseManager(
                     elif candidate_type == 'text' and ('续表' in candidate_title or '续表' in candidate_content):
                         # 遇到"续表"文本，继续查找下一个表格
                         continue
-                    elif candidate_type not in ('text',):
+                    elif candidate_type == 'text':
+                        # 非"续表"文本，收集后停止查找
+                        intermediate_texts.append(candidate)
+                        break
+                    else:
                         # 遇到非文本类型，停止查找
                         break
 
@@ -439,12 +489,22 @@ class KnowledgeBaseManager(
                     # 获取内容（用于检测"续表"）
                     next_content = getattr(next_chunk, 'content', '')
 
+                    # 通用/无意义标题集合，这些标题不能用于"标题相似"判定
+                    _GENERIC_TITLES = {'表格', 'table', '表格', ''}
+
                     # 判断是否为跨页表格
                     is_cross_page = False
 
                     # 规则1: 页码连续（如果页码有效）
                     page_valid = curr_page_end > 0 and next_page_start > 0
                     if page_valid and curr_page_end + 1 == next_page_start:
+                        # 页码连续时，还需标题匹配或为通用标题才合并
+                        # 避免把不同页面上不相关的表格错误合并
+                        if curr_title == next_title or curr_title in _GENERIC_TITLES and next_title in _GENERIC_TITLES:
+                            is_cross_page = True
+                        elif curr_title and next_title:
+                            clean_next_r1 = next_title.replace('续表', '').strip()
+                            if curr_title in clean_next_r1 or clean_next_r1 in curr_title:
                                 is_cross_page = True
 
                     # 规则2: 第二个表格标题或内容包含"续表"
@@ -452,11 +512,36 @@ class KnowledgeBaseManager(
                         is_cross_page = True
 
                     # 规则3: 标题相似（去掉"续表"后比较）
-                    elif curr_title and next_title:
+                    # 排除通用标题（如"表格"），防止把所有标题为"表格"的相邻表格都误合并
+                    elif (curr_title and next_title
+                          and curr_title not in _GENERIC_TITLES
+                          and next_title not in _GENERIC_TITLES):
                         clean_next = next_title.replace('续表', '').strip()
-                        if curr_title in clean_next or clean_next in curr_title:
+                        if clean_next and (curr_title in clean_next or clean_next in curr_title):
                             is_cross_page = True
 
+                    # 规则4（降级）: 页码不可用（如 Word 文档 page_idx 全为 0）
+                    # 仅当页码信息缺失时才启用此规则，避免 PDF 正常页码时被误合并
+                    if (not is_cross_page
+                            and pages_unavailable
+                            and curr_title in _GENERIC_TITLES
+                            and next_title in _GENERIC_TITLES):
+                        # 检查中间文本是否暗示跨页延续（空、短文本、续表标记等）
+                        has_separating_content = False
+                        for text_chunk in intermediate_texts:
+                            tc = (getattr(text_chunk, 'content', '') or '').strip()
+                            tt = (getattr(text_chunk, 'title', '') or '').strip()
+                            if not tc:
+                                continue  # 空文本不算分隔
+                            if '续表' in tc or '续表' in tt:
+                                continue  # 续表标记，说明是跨页
+                            # 有实质性中间内容（如分类标题"A3类：xxx"），不合并
+                            has_separating_content = True
+                            break
+                        if not has_separating_content:
+                            is_cross_page = True
+                            logger.debug(f"降级合并(页码不可用): '{curr_title}' + '{next_title}'")
+
                     if is_cross_page:
                         # 执行合并
                         merge_count += 1
@@ -466,17 +551,61 @@ class KnowledgeBaseManager(
                         curr_html = getattr(current, 'table_html', '') or ''
                         next_html = getattr(next_chunk, 'table_html', '') or ''
                         if curr_html and next_html:
-                            # 合并两个表格的 HTML
+                            # 正确合并两个表格的 HTML：
+                            # 将第二个表格的 <tr> 行追加到第一个表格中
+                            # （而非简单拼接两个 <table>，否则 html_table_to_markdown
+                            #  的 soup.find('table') 只能找到第一个表格）
+                            try:
+                                soup1 = BeautifulSoup(curr_html, 'html.parser')
+                                soup2 = BeautifulSoup(next_html, 'html.parser')
+                                table1 = soup1.find('table')
+                                table2 = soup2.find('table')
+                                if table1 and table2:
+                                    # 从第二个表格提取数据行
+                                    next_rows = table2.find_all('tr')
+                                    # 跳过与第一个表格表头重复的行
+                                    # 对比第一行而非所有 th（find_all('th') 会匹配
+                                    # 整个表格的 th，无法与单行做列表比较）
+                                    first_row_t1 = table1.find('tr')
+                                    if first_row_t1 and next_rows:
+                                        row1_texts = [c.get_text(strip=True) for c in first_row_t1.find_all(['th', 'td'])]
+                                        row2_texts = [c.get_text(strip=True) for c in next_rows[0].find_all(['th', 'td'])]
+                                        if row1_texts and row2_texts and row1_texts == row2_texts:
+                                            next_rows = next_rows[1:]
+                                            logger.debug("跨页表格合并: 跳过了重复的表头行")
+                                    for row in next_rows:
+                                        table1.append(row)
+                                    current.table_html = str(soup1)
+                                    logger.debug(f"跨页表格 HTML 合并成功: 追加了 {len(next_rows)} 行")
+                                else:
                                     current.table_html = curr_html + '\n' + next_html
+                            except Exception as e:
+                                logger.warning(f"跨页表格 HTML 合并异常: {e}，回退到简单拼接")
+                                current.table_html = curr_html + '\n' + next_html
+                        elif not curr_html and next_html:
+                            current.table_html = next_html
 
-                        # 合并 image_path 到 images
+                        # 合并 image_path 和嵌入图片到 images
                         curr_img = getattr(current, 'image_path', None)
                         next_img = getattr(next_chunk, 'image_path', None)
-                        merged_images = []
-                        if curr_img:
+                        curr_images = getattr(current, 'images', None) or []
+                        next_images = getattr(next_chunk, 'images', None) or []
+
+                        # 合并两个表格的所有图片（image_path + 嵌入图片）
+                        merged_images = list(curr_images)  # 保留当前表格的嵌入图片
+                        # 添加 image_path 图片（如果不在列表中）
+                        existing_ids = {img.get('id', '') for img in merged_images if isinstance(img, dict)}
+                        if curr_img and curr_img not in existing_ids:
                             merged_images.append({'id': curr_img, 'page': curr_page_end})
-                        if next_img:
+                            existing_ids.add(curr_img)
+                        for img in next_images:  # 添加下一个表格的嵌入图片
+                            img_id = img.get('id', '') if isinstance(img, dict) else ''
+                            if img_id and img_id not in existing_ids:
+                                merged_images.append(img)
+                                existing_ids.add(img_id)
+                        if next_img and next_img not in existing_ids:
                             merged_images.append({'id': next_img, 'page': next_page_start})
+
                         if merged_images:
                             current.images = merged_images
                             # 保留第一个图片作为主 image_path
@@ -525,7 +654,7 @@ class KnowledgeBaseManager(
         try:
             from config import get_llm_client, DASHSCOPE_MODEL
             client = get_llm_client()
-            summary = call_llm(client, prompt, DASHSCOPE_MODEL, max_tokens=100)
+            summary = call_llm(client, prompt, DASHSCOPE_MODEL, max_tokens=2048)
             return summary.strip() if summary else ""
         except Exception as e:
             logger.warning(f"生成表格摘要失败: {e}")
@@ -584,13 +713,31 @@ class KnowledgeBaseManager(
                         ]
                     }
                 ],
-                max_tokens=200
+                max_tokens=2048  # mimo-v2.5 推理模型思考链消耗 ~1000 token，需留足输出空间
             )
 
             description = response.choices[0].message.content
 
+            # 推理模型兼容：content 为空时从 reasoning_content 提取
+            if not description or not description.strip():
+                reasoning = getattr(response.choices[0].message, 'reasoning_content', None)
+                if reasoning and reasoning.strip():
+                    import re
+                    # 尝试从思考链中提取有用文本（去掉 <think> 标签后的内容）
+                    cleaned = re.sub(r'', '', reasoning, flags=re.DOTALL).strip()
+                    if cleaned:
+                        logger.info(f"VLM content为空，从reasoning_content提取描述: {image_path}")
+                        description = cleaned
+                    else:
+                        description = reasoning.strip()
+
+            if not description:
+                logger.warning(f"VLM 返回空描述: {image_path}")
+                return ""
+
             # 缓存结果
             import hashlib
+            import re as _re
             img_hash = hashlib.md5(img_path.read_bytes()).hexdigest()
             cache_dir = Path('.data/cache/vlm')
             cache_dir.mkdir(parents=True, exist_ok=True)

knowledge/router.py

@@ -283,7 +283,7 @@ class KnowledgeBaseRouter:
         content = call_llm(
             self.llm_client, prompt, MODEL,
             temperature=0.1,
-            max_tokens=100
+            max_tokens=512
         )
 
         if content is None:

knowledge/sync.py

@@ -664,6 +664,17 @@ class KnowledgeSyncService:
                 except Exception as e:
                     logger.warning(f"递增缓存版本号失败: {e}")
 
+            # 语义缓存无版本号机制，文档变更后必须清空，
+            # 否则可能返回过时的 images/sources/citations（如已删除的图片 404）
+            try:
+                from core.semantic_cache import get_semantic_cache
+                _sc = get_semantic_cache()
+                if _sc:
+                    _sc.clear()
+                    logger.debug(f"已清空语义缓存（文档变更触发）: {kb_name}")
+            except Exception as e:
+                logger.warning(f"清空语义缓存失败: {e}")
+
             return True
 
         except Exception as e:

parsers/heading_rules.py

@@ -0,0 +1,374 @@
+# -*- coding: utf-8 -*-
+"""
+标题识别规则引擎
+
+将 _detect_heading_level 的硬编码正则提取为可配置的规则列表。
+规则按优先级从高到低排序，第一个匹配即返回。
+
+MinerU 解析 DOCX 等 Office 格式时通常不提供 text_level（全部为 0），
+此时需要启发式识别标题层级。本模块提供可配置的规则引擎替代原来的
+硬编码 if-elif 链。
+
+设计要点：
+- HeadingRule 数据类支持正向匹配（pattern）和反向排除（exclude_pattern）
+- 长度约束（min_length / max_length）可精确控制匹配范围
+- 规则可单独禁用（enabled=False），便于调试
+- 全局单例通过 config.py 覆盖默认值
+
+MinerU v2 格式备注：
+    content_list_v2.json 中的 paragraph_content 包含 style=["bold"] 信息，
+    layout.json 中的 spans 也有 style 信息。这些信息比正则匹配 **加粗** 更可靠，
+    但当前代码使用 v1 格式（content_list.json），暂不利用 v2 的 style。
+    HeadingRuleEngine.detect() 签名预留了 style 参数，未来切换到 v2 格式后
+    可直接利用 style 信息辅助判断。
+"""
+
+import re
+import logging
+from dataclasses import dataclass
+from typing import Optional, List, Tuple
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class HeadingRule:
+    """
+    标题识别规则
+
+    每条规则定义一个文本模式到标题级别的映射。
+    规则引擎按列表顺序逐条匹配，第一个命中即返回。
+
+    Attributes:
+        pattern: 编译后的正则（match 语义，从文本开头匹配）
+        level: 匹配时返回的标题级别 (1=h1, 2=h2, 3=h3)
+        name: 规则名称（用于日志和配置覆盖）
+        max_length: 文本最大长度，0=不限
+        min_length: 文本最小长度，0=不限
+        enabled: 是否启用
+        exclude_pattern: 匹配此模式则排除（反向过滤）
+
+    Example:
+        >>> rule = HeadingRule(
+        ...     pattern=re.compile(r'^第[一二三四五六七八九十百千万]+[章节篇部]'),
+        ...     level=1,
+        ...     name="chinese_chapter",
+        ... )
+        >>> rule.match("第一章 总则")
+        1
+        >>> rule.match("这是正文")
+        0
+    """
+
+    pattern: re.Pattern
+    level: int
+    name: str
+    max_length: int = 0
+    min_length: int = 0
+    enabled: bool = True
+    exclude_pattern: Optional[re.Pattern] = None
+
+    def match(self, text: str) -> int:
+        """
+        检查文本是否匹配此规则
+
+        Args:
+            text: 待检测文本（调用前应已 strip）
+
+        Returns:
+            标题级别，0 表示不匹配
+        """
+        if not self.enabled:
+            return 0
+        if self.min_length > 0 and len(text) < self.min_length:
+            return 0
+        if self.max_length > 0 and len(text) > self.max_length:
+            return 0
+        if self.exclude_pattern and self.exclude_pattern.search(text):
+            return 0
+        if self.pattern.match(text):
+            return self.level
+        return 0
+
+
+# 默认规则列表（按优先级从高到低）
+#
+# 注意事项：
+# - 数字三级标题 (1.1.1) 必须在二级 (1.1) 之前，因为 1.1.1 也匹配 ^\d+\.\d+
+# - short_chinese_heading 是最宽泛的规则，放在最后作为兜底
+# - 第 9 条规则相比原版增加了 exclude_pattern，排除以句末标点结尾的短文本
+DEFAULT_HEADING_RULES: List[HeadingRule] = [
+    # 1. 中文章节标题 -> h1
+    #    匹配：第一章、第二章、第十节、第三篇 等
+    HeadingRule(
+        pattern=re.compile(r'^第[一二三四五六七八九十百千万]+[章节篇部]'),
+        level=1,
+        name="chinese_chapter",
+    ),
+    # 2. 中文条款编号 -> h2
+    #    匹配：第一条、第三款 等（仅短标题，长正文段落不算标题）
+    HeadingRule(
+        pattern=re.compile(r'^第[一二三四五六七八九十百千万]+[条款]'),
+        level=2,
+        name="chinese_article",
+        max_length=30,
+    ),
+    # 3. 数字三级标题 -> h3（必须在二级之前匹配）
+    #    匹配：1.1.1 背景、2.3.4 方案 等
+    HeadingRule(
+        pattern=re.compile(r'^\d+\.\d+\.\d+[\.、\s]'),
+        level=3,
+        name="numeric_level3",
+        max_length=100,
+    ),
+    # 4. 数字二级标题 -> h2（必须在一级之前匹配）
+    #    匹配：1.1 背景、2.3 方案、2.1运行调度（无空格） 等
+    #    使用负向前瞻排除三级标题（由 numeric_level3 处理）
+    HeadingRule(
+        pattern=re.compile(r'^\d+\.\d+(?!\.\d)'),
+        level=2,
+        name="numeric_level2",
+        max_length=80,
+    ),
+    # 5. 数字一级标题 -> h1
+    #    匹配：1. 概述、2、背景 等
+    #    排除：以 ；;。，、： 结尾的文本（这些是编号列表项/子条目，不是独立标题）
+    HeadingRule(
+        pattern=re.compile(r'^\d+[\.、\s]'),
+        level=1,
+        name="numeric_level1",
+        max_length=50,
+        exclude_pattern=re.compile(r'[；;。，、：:]$'),
+    ),
+    # 6. 英文章节标题 -> h1
+    #    匹配：Chapter 1、Section 2、Part 3 等
+    HeadingRule(
+        pattern=re.compile(r'^(Chapter|Section|Part|Chapter\s+\d+|Section\s+\d+)', re.IGNORECASE),
+        level=1,
+        name="english_chapter",
+    ),
+    # 7. 分类标题 -> h3（必须在 bold_short_text 之前，否则 **A2类：** 会被加粗规则抢先匹配）
+    #    匹配：A1类：公园、**A2类**：各类卫生医疗机构、**B1类：** 道路 等
+    HeadingRule(
+        pattern=re.compile(r'^\*{0,2}[A-Z]\d+[类類]\*{0,2}[：:]'),
+        level=3,
+        name="category_heading",
+    ),
+    # 8. 加粗短文本 -> h2
+    #    匹配：**重要通知**、**概述** 等（Markdown 加粗标记）
+    #    注意：**A2类：** 已被分类标题规则优先匹配，不会误判为 h2
+    HeadingRule(
+        pattern=re.compile(r'^\*\*.+\*\*$'),
+        level=2,
+        name="bold_short_text",
+        max_length=50,
+    ),
+    # 9. 短中文文本 -> h2（替代原"任何 <20 字符含中文"规则）
+    #    关键改进：排除以句末标点结尾的文本
+    #    原规则将 "这是一段正文。" 也识别为 h2，导致大量误判
+    #    新规则：包含中文 + 长度 2-20 + 不以句末标点结尾 → h2
+    HeadingRule(
+        pattern=re.compile(r'[一-鿿]'),
+        level=2,
+        name="short_chinese_heading",
+        max_length=20,
+        min_length=2,
+        exclude_pattern=re.compile(r'[。！？；…]$'),
+        enabled=True,
+    ),
+]
+
+
+class HeadingRuleEngine:
+    """
+    标题识别规则引擎
+
+    按规则列表顺序逐条匹配，第一个命中即返回标题级别。
+    支持从 config.py 加载自定义规则或覆盖默认规则参数。
+
+    Example:
+        >>> engine = HeadingRuleEngine()
+        >>> engine.detect("第一章 总则")
+        (1, 'chinese_chapter')
+        >>> engine.detect("这是普通正文。")
+        (0, None)
+    """
+
+    def __init__(self, rules: Optional[List[HeadingRule]] = None) -> None:
+        """
+        Args:
+            rules: 规则列表，None 则使用默认规则的深拷贝
+        """
+        if rules is not None:
+            self.rules: List[HeadingRule] = rules
+        else:
+            import copy
+            self.rules = copy.deepcopy(DEFAULT_HEADING_RULES)
+
+    def _validate_level(self, level: int, text: str, rule_name=None):
+        """各级别标题长度防护：超长文本不应作为标题，降为正文。
+        H1 > 40字, H2 > 60字, H3 > 50字 → 降为正文。
+        统一覆盖 v1 常规匹配、v2 style 匹配、bold_short_text 兜底所有返回路径。"""
+        text_len = len(text)
+        if level == 1 and text_len > 40:
+            logger.debug(f"标题识别: '{text[:30]}...' H1 但超长({text_len}字)，降为正文")
+            return 0, None
+        if level == 2 and text_len > 60:
+            logger.debug(f"标题识别: '{text[:30]}...' H2 但超长({text_len}字)，降为正文")
+            return 0, None
+        if level == 3 and text_len > 50:
+            logger.debug(f"标题识别: '{text[:30]}...' H3 但超长({text_len}字)，降为正文")
+            return 0, None
+        return level, rule_name
+
+    def detect(self, text: str, style: Optional[List[str]] = None) -> Tuple[int, Optional[str]]:
+        """
+        检测文本的标题级别
+
+        Args:
+            text: 待检测文本
+            style: MinerU v2 格式中的 style 信息（如 ["bold"]），
+                   当文本标记为 bold 且较短时，可直接判定为标题，
+                   无需依赖 Markdown **...** 标记。
+
+        Returns:
+            (level, rule_name): 标题级别和匹配的规则名
+            level=0 表示不是标题
+        """
+        text = text.strip()
+        if not text:
+            return 0, None
+
+        # v2 style 信息：如果文本标记为 bold 且较短，优先尝试加粗规则
+        if style and 'bold' in style and 2 <= len(text) <= 50:
+            # 先检查是否匹配更高优先级的分类标题规则
+            for rule in self.rules:
+                if rule.name == 'category_heading' and rule.enabled:
+                    level = rule.match(text)
+                    if level > 0:
+                        logger.debug(f"标题识别(v2 style): '{text[:30]}' -> h{level} (规则: {rule.name})")
+                        return self._validate_level(level, text, rule.name)
+
+            # 再检查是否匹配中文章节/条款等高优先级规则
+            for rule in self.rules:
+                if rule.name in ('chinese_chapter', 'chinese_article', 'numeric_level3',
+                                 'numeric_level2', 'numeric_level1', 'english_chapter') and rule.enabled:
+                    level = rule.match(text)
+                    if level > 0:
+                        logger.debug(f"标题识别(v2 style): '{text[:30]}' -> h{level} (规则: {rule.name})")
+                        return self._validate_level(level, text, rule.name)
+
+            # 最后兜底：加粗短文本 → h2
+            # 但需先检查所有规则的 exclude_pattern，防止编号列表项被误判为标题
+            # 例如 "3.完全满足品规。指..." 虽有 bold 样式，但属于列表项而非标题
+            for rule in self.rules:
+                if rule.enabled and rule.exclude_pattern and rule.exclude_pattern.search(text):
+                    logger.debug(f"标题识别(v2 style): '{text[:30]}' 被 {rule.name} 的 exclude_pattern 排除")
+                    return 0, None
+
+            logger.debug(f"标题识别(v2 style): '{text[:30]}' -> h2 (规则: bold_short_text_via_style)")
+            return 2, 'bold_short_text'
+
+        # 常规规则匹配（v1 格式或无 style 信息时）
+        for rule in self.rules:
+            level = rule.match(text)
+            if level > 0:
+                logger.debug(f"标题识别: '{text[:30]}' -> h{level} (规则: {rule.name})")
+                return self._validate_level(level, text, rule.name)
+
+        return 0, None
+
+
+# ==================== 全局单例 ====================
+
+_engine: Optional[HeadingRuleEngine] = None
+
+
+def get_heading_engine() -> HeadingRuleEngine:
+    """获取全局标题识别引擎（延迟初始化，线程安全）"""
+    global _engine
+    if _engine is None:
+        _engine = _create_engine_from_config()
+    return _engine
+
+
+def _create_engine_from_config() -> HeadingRuleEngine:
+    """
+    从 config 创建引擎（支持配置覆盖）
+
+    优先级：
+    1. config.HEADING_RULES_CONFIG 不为 None → 使用自定义规则
+    2. config 细粒度参数覆盖默认规则（如 HEADING_SHORT_TEXT_ENABLED）
+    3. 使用默认规则
+    """
+    # 尝试加载完整自定义规则
+    try:
+        from config import HEADING_RULES_CONFIG
+        if HEADING_RULES_CONFIG is not None:
+            rules = _build_rules_from_config(HEADING_RULES_CONFIG)
+            logger.info(f"使用自定义标题规则: {len(rules)} 条")
+            return HeadingRuleEngine(rules)
+    except ImportError:
+        pass
+
+    # 使用默认规则，应用细粒度配置覆盖
+    rules = list(DEFAULT_HEADING_RULES)
+    try:
+        from config import HEADING_SHORT_TEXT_ENABLED
+        for rule in rules:
+            if rule.name == "short_chinese_heading":
+                rule.enabled = HEADING_SHORT_TEXT_ENABLED
+                logger.debug(f"配置覆盖: short_chinese_heading.enabled={HEADING_SHORT_TEXT_ENABLED}")
+    except ImportError:
+        pass
+
+    try:
+        from config import HEADING_SHORT_TEXT_MAX_LENGTH
+        for rule in rules:
+            if rule.name == "short_chinese_heading":
+                rule.max_length = HEADING_SHORT_TEXT_MAX_LENGTH
+                logger.debug(f"配置覆盖: short_chinese_heading.max_length={HEADING_SHORT_TEXT_MAX_LENGTH}")
+    except ImportError:
+        pass
+
+    return HeadingRuleEngine(rules)
+
+
+def _build_rules_from_config(config: list) -> List[HeadingRule]:
+    """
+    从配置字典列表构建规则列表
+
+    Args:
+        config: 规则配置列表，每项为 dict，包含：
+            - pattern (str): 正则表达式字符串
+            - level (int): 标题级别
+            - name (str): 规则名称
+            - max_length (int, 可选): 文本最大长度
+            - min_length (int, 可选): 文本最小长度
+            - enabled (bool, 可选): 是否启用
+            - exclude_pattern (str, 可选): 排除正则
+
+    Returns:
+        规则列表
+    """
+    rules = []
+    for item in config:
+        exclude = None
+        if 'exclude_pattern' in item:
+            exclude = re.compile(item['exclude_pattern'])
+        rules.append(HeadingRule(
+            pattern=re.compile(item['pattern']),
+            level=item['level'],
+            name=item['name'],
+            max_length=item.get('max_length', 0),
+            min_length=item.get('min_length', 0),
+            enabled=item.get('enabled', True),
+            exclude_pattern=exclude,
+        ))
+    return rules
+
+
+def reset_heading_engine() -> None:
+    """重置引擎（用于测试）"""
+    global _engine
+    _engine = None

services/feedback.py

@@ -616,7 +616,7 @@ class FeedbackService:
                 prompt=prompt,
                 model=self.model,
                 temperature=0.7,
-                max_tokens=200
+                max_tokens=512
             )
 
             if not response:

services/session.py

@@ -159,7 +159,7 @@ class SessionManager:
                 SELECT id, role, content, metadata, created_at
                 FROM messages
                 WHERE session_id = ?
-                ORDER BY created_at DESC
+                ORDER BY created_at DESC, id DESC
                 LIMIT ?
             ''', (session_id, limit))