sync(server-release): 从 main 同步逻辑改动（不含 API 格式变更）

同步内容:
- knowledge/manager.py: VLM max_tokens 512→2048 + reasoning_content fallback
- services/feedback.py: FAQ max_tokens 200→512
- services/session.py: 消息排序增加 id DESC 次级排序
- knowledge/image_cleanup.py: 新增图片/VLM缓存孤儿文件清理模块
- knowledge/collection.py: 集合删除时清理孤儿文件 + list_collections 磁盘扫描策略
- knowledge/document.py: 文档删除时清理孤儿文件
- api/chat_routes.py: 新增 _rescue_bm25_divergence 函数（BM25-CE分歧救援）
- core/intent_analyzer.py: 使用 get_intent_client 专用客户端 + 移除短追问缓存跳过逻辑
- cleanup_orphans.py: 独立孤儿文件清理脚本

knowledge/collection.py

@@ -331,6 +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()
@@ -361,8 +376,40 @@ class CollectionMixin:
         self._metadata = self._load_metadata()
         result = []
 
-        # 以元数据为唯一真相源，不再扫描磁盘自动补充
-        # （扫描磁盘会导致其他 worker 刚创建/删除的集合被误操作）
+        # 扫描 base_path 下的所有子目录作为向量库
+        # 每个向量库使用独立目录：base_path/my_ky, base_path/public_kb 等
+        actual_collections = []
+        try:
+            if os.path.exists(self.base_path):
+                for item in os.listdir(self.base_path):
+                    item_path = os.path.join(self.base_path, item)
+                    if os.path.isdir(item_path) and not item.startswith('.'):
+                        # 检查是否包含 chroma.sqlite3（有效的向量库目录）
+                        if os.path.exists(os.path.join(item_path, 'chroma.sqlite3')):
+                            actual_collections.append(item)
+        except Exception as e:
+            logger.warning(f"扫描向量库目录失败: {e}")
+
+        # 如果扫描失败，回退到元数据中的集合列表
+        if not actual_collections:
+            actual_collections = list(self._metadata.get("collections", {}).keys())
+
+        for name in actual_collections:
+            if name not in self._metadata.get("collections", {}):
+                if "collections" not in self._metadata:
+                    self._metadata["collections"] = {}
+                self._metadata["collections"][name] = {
+                    "display_name": name,
+                    "department": "",
+                    "description": "",
+                    "created_at": datetime.now().isoformat()
+                }
+                logger.info(f"自动补充向量库元数据: {name}")
+
+        self._save_metadata()
+
+        stale_collections = []
+
         for name, info in self._metadata.get("collections", {}).items():
             try:
                 collection = self.get_collection(name)
@@ -375,18 +422,17 @@ class CollectionMixin:
                     description=info.get("description", "")
                 ))
             except Exception as e:
-                # 只跳过不修改元数据（可能是其他 worker 刚创建的集合）
                 logger.warning(
-                    f"跳过异常向量库 '{name}': {e}"
+                    f"跳过异常向量库 '{name}': {e}，可能是 ChromaDB 数据目录已丢失"
                 )
-                result.append(CollectionInfo(
-                    name=name,
-                    display_name=info.get("display_name", name),
-                    document_count=0,
-                    created_at=info.get("created_at", ""),
-                    department=info.get("department", ""),
-                    description=info.get("description", "")
-                ))
+                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
 

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/manager.py

@@ -654,7 +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=512)
+            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}")
@@ -713,13 +713,31 @@ class KnowledgeBaseManager(
                         ]
                     }
                 ],
-                max_tokens=512
+                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)