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

api/chat_routes.py

@@ -293,6 +293,112 @@ def _process_table_doc(doc: str, meta: Dict) -> str:
     return doc
 
 
+def _rescue_bm25_divergence(contexts: List[Dict], search_result: dict,
+                            min_score: float) -> List[Dict]:
+    """
+    BM25-CrossEncoder 分歧检测救援：当 BM25 排名靠前（top-3）的切片
+    被 CrossEncoder rerank 压制（分数低于 min_score 或被截断）时，
+    将其分数提升至保底值，使其通过后续的 min_score 过滤。
+
+    适用场景：
+    - BM25 top-1/top-2 的切片（精确关键词匹配强信号）被 rerank 评分极低
+    - 正确切片在 rerank 后被截断，根本不在 contexts 中
+
+    Args:
+        contexts: 全部上下文切片
+        search_result: engine.search_hybrid() 返回的原始结果（含 _bm25_top3）
+        min_score: 最低分数阈值
+
+    Returns:
+        修改后的 contexts
+    """
+    if not contexts:
+        return contexts
+
+    from config import (
+        BM25_DIVERGENCE_RESCUE_ENABLED, BM25_DIVERGENCE_MAX_RANK,
+        CLUSTER_RESCUE_FLOOR,
+    )
+
+    if not BM25_DIVERGENCE_RESCUE_ENABLED:
+        return contexts
+
+    bm25_top3 = search_result.get('_bm25_top3', [])
+    if not bm25_top3:
+        return contexts
+
+    # 构建 contexts 中已有切片的 ID 索引，用于快速查找
+    existing_ids = {}
+    for i, ctx in enumerate(contexts):
+        chunk_id = ctx.get('meta', {}).get('chunk_id') or ctx.get('id')
+        if chunk_id:
+            existing_ids[chunk_id] = i
+
+    # 计算 contexts 中的多数 source（用于情况 B 注入校验）
+    # 防止跨文档注入无关切片（如 q013 场景：2.docx 的吸烟场所切片被注入到 1.docx 的查询中）
+    source_counter: Dict[str, int] = {}
+    for ctx in contexts:
+        src = ctx.get('meta', {}).get('source', '')
+        if src:
+            source_counter[src] = source_counter.get(src, 0) + 1
+    majority_source = max(source_counter, key=source_counter.get) if source_counter else None
+
+    rescued_count = 0
+
+    for bm25_item in bm25_top3:
+        rank = bm25_item.get('rank', 99)
+        if rank > BM25_DIVERGENCE_MAX_RANK:
+            continue
+
+        bm25_id = bm25_item.get('id')
+        bm25_meta = bm25_item.get('meta', {})
+        bm25_doc = bm25_item.get('doc', '')
+
+        # 情况 A：切片在 contexts 中但 score < min_score
+        if bm25_id and bm25_id in existing_ids:
+            ctx = contexts[existing_ids[bm25_id]]
+            if ctx.get('score', 0) < min_score:
+                ctx['score'] = CLUSTER_RESCUE_FLOOR
+                rescued_count += 1
+                logger.debug(
+                    f"BM25 分歧救援 (情况A): rank={rank}, "
+                    f"source={bm25_meta.get('source', '')}, "
+                    f"section={bm25_meta.get('section', '')}, "
+                    f"原score→{CLUSTER_RESCUE_FLOOR}"
+                )
+            continue
+
+        # 情况 B：切片不在 contexts 中（被 rerank 截断或已被过滤）
+        # 从 _bm25_top3 备份中注入，但需校验 source 一致性
+        if bm25_doc and bm25_meta:
+            bm25_source = bm25_meta.get('source', '')
+            # source 一致性校验：只注入与 contexts 多数 source 一致的切片
+            if majority_source and bm25_source and bm25_source != majority_source:
+                logger.debug(
+                    f"BM25 分歧救援 (情况B-跳过): rank={rank}, "
+                    f"source={bm25_source} != majority={majority_source}, "
+                    f"section={bm25_meta.get('section', '')}"
+                )
+                continue
+            injected_ctx = {
+                'doc': bm25_doc,
+                'meta': bm25_meta,
+                'score': CLUSTER_RESCUE_FLOOR,
+            }
+            contexts.append(injected_ctx)
+            rescued_count += 1
+            logger.debug(
+                f"BM25 分歧救援 (情况B-注入): rank={rank}, "
+                f"source={bm25_meta.get('source', '')}, "
+                f"section={bm25_meta.get('section', '')}"
+            )
+
+    if rescued_count > 0:
+        logger.info(f"BM25 分歧救援: 共救援 {rescued_count} 个切片")
+
+    return contexts
+
+
 def _rescue_lexical_match(contexts: List[Dict], retrieval_query: str,
                           min_score: float) -> List[Dict]:
     """
@@ -2518,6 +2624,9 @@ def rag():
             # 4. 构建 prompt（Phase 6：LLM 图片感知）
             # Bug 1 修复：文本切片用于 top 5 名额竞争，图片描述不参与竞争
 
+            # BM25 分歧检测救援：BM25 top-3 但 rerank 压制的切片
+            contexts = _rescue_bm25_divergence(contexts, search_result, RERANK_CONTEXT_MIN_SCORE)
+
             # 词法匹配救援：当切片文本精确包含查询关键词但 CrossEncoder 评分低时，
             # 提升分数使其通过 min_score 过滤（适用于独立切片无法触发聚类救援的场景）
             contexts = _rescue_lexical_match(contexts, retrieval_query, RERANK_CONTEXT_MIN_SCORE)

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()

core/intent_analyzer.py

@@ -233,10 +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
+            from config import get_intent_client
-            self._client = get_llm_client()
+            self._client = get_intent_client()
         return self._client
 
     def _get_cache(self):
@@ -309,9 +309,7 @@ class IntentAnalyzer:
 
         # 2. 尝试从语义缓存获取
         cache = self._get_cache()
-        # 短追问（如"B3类呢？"）严重依赖上下文，embedding 区分度不足易误命中
+        if cache:
-        is_short_followup = history and len(query.strip()) < 20
-        if cache and not is_short_followup:
             # 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)
@@ -327,10 +325,7 @@ class IntentAnalyzer:
             else:
                 logger.warning("意图分析缓存: _get_embedding 返回 None，跳过缓存")
         else:
-            if is_short_followup:
+            logger.warning("意图分析缓存: _get_cache 返回 None，缓存不可用")
-                logger.debug(f"跳过语义缓存(短追问依赖上下文): query='{query[:30]}'")
-            elif not cache:
-                logger.warning("意图分析缓存: _get_cache 返回 None，缓存不可用")
 
         # 构建 prompt
         user_prompt = f"""## 对话历史

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 base_path 下的所有子目录作为向量库
-        # （扫描磁盘会导致其他 worker 刚创建/删除的集合被误操作）
+        # 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(
+                stale_collections.append(name)
-                    name=name,
+
-                    display_name=info.get("display_name", name),
+        # 清理元数据中指向已失效集合的条目
-                    document_count=0,
+        if stale_collections:
-                    created_at=info.get("created_at", ""),
+            for name in stale_collections:
-                    department=info.get("department", ""),
+                self._metadata.get("collections", {}).pop(name, None)
-                    description=info.get("description", "")
+                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)

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))