feat(rag): 子章节级图片过滤 + VLM 后台增强 + 意图分析优化

- 图片选择新增 section_path 字段，支持子章节级过滤
- _filter_images_by_answer 增加 primary_sections 参数和叶子节点匹配
- 检索发散检测：primary_leaf_names > 3 时全局阈值+1
- _FIGURE_ANSWER_KEYWORDS 移除单字"图""表"，正则图号兜底防误触发
- lazy_enhance 后台增强流程优化
- 意图分析与 LLM 工具层改进

评测：图片选择 F1 从 52.4% 提升至 66.3%（+13.9pp），Precision +16.5pp

knowledge/manager.py

@@ -622,7 +622,7 @@ 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}")
@@ -681,13 +681,31 @@ 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)