feat: 性能插桩、缓存清除端点、LLM top_p 与 JSON 强制输出

- chat_routes.py: 新增 8 阶段性能计时（stages_ms）+ SSE 输出
- sync_routes.py: 新增 POST /cache/clear 缓存清除端点
- engine.py: LLM 调用增加 top_p 参数
- intent_analyzer.py: response_format 加配置开关（INTENT_RESPONSE_FORMAT）

api/chat_routes.py

@@ -2196,6 +2196,9 @@ def rag():
         import re
         start_time = _time.time()
         full_answer = []
+        # === 性能插桩：各阶段耗时 ===
+        _t = {'intent': 0, 'cache': 0, 'search': 0, 'vlm': 0, 'image': 0, 'rescue': 0, 'llm': 0, 'post': 0}
+        _t_prev = start_time
 
         try:
             # 0. 意图分析（改写 + 双层判断）
@@ -2227,7 +2230,6 @@ def rag():
                     # 构建上下文
                     context_text = ""
                     if history:
-                        # 提取最近的助手回答
                         for msg in reversed(history):
                             if msg.get("role") == "assistant":
                                 context_text = msg.get("content", "")
@@ -2260,7 +2262,6 @@ def rag():
 
 请直接回答用户问题。"""
 
-                    # 流式生成回答
                     for content in call_llm_stream(
                         client,
                         prompt=user_prompt,
@@ -2274,14 +2275,15 @@ def rag():
                         full_answer.append(content)
                         yield f"data: {json.dumps({'type': 'chunk', 'content': content}, ensure_ascii=False)}\n\n"
 
-                    # 发送完成事件
                     yield f"data: {json.dumps({'type': 'finish', 'answer': ''.join(full_answer), 'sources': []}, ensure_ascii=False)}\n\n"
-                    return  # 直接返回，不执行后续检索
+                    return
 
             except Exception as e:
                 logger.warning(f"意图分析失败: {e}，继续执行检索流程")
 
             retrieval_query = intent.rewritten_query if (intent and intent.rewritten_query) else message
+            _t['intent'] = round((_time.time() - _t_prev) * 1000, 1)
+            _t_prev = _time.time()
 
             # 1.5 语义缓存检查（跳过检索+生成全流程）
             _semantic_cache_emb = None
@@ -2343,6 +2345,8 @@ def rag():
 
             # 1. 发送开始事件
             yield f"data: {json.dumps({'type': 'start', 'message': '正在检索知识库...'}, ensure_ascii=False)}\n\n"
+            _t['cache'] = round((_time.time() - _t_prev) * 1000, 1)
+            _t_prev = _time.time()
 
             # 2. 执行混合检索（扩大召回数量，确保图片切片有机会被召回）
             # 如果意图分析生成了子查询（对比类），传给搜索引擎并行检索
@@ -2619,6 +2623,8 @@ def rag():
             # 2.5. 懒加载增强（Phase 4）
             # 暂时禁用：VLM 调用耗时过长，可能导致请求超时
             # TODO: 后续可改为异步后台任务
+            _t['search'] = round((_time.time() - _t_prev) * 1000, 1)
+            _t_prev = _time.time()
             try:
                 import asyncio
                 from knowledge.lazy_enhance import enhance_retrieved_chunks
@@ -2626,9 +2632,13 @@ def rag():
                 asyncio.run(enhance_retrieved_chunks(contexts, message, kb_name))
             except Exception as e:
                 logger.warning(f"懒加载增强失败: {e}")
+            _t['vlm'] = round((_time.time() - _t_prev) * 1000, 1)
+            _t_prev = _time.time()
 
             # 3. 选择要展示的图片（Phase 5）
             selected_images = select_images(contexts, message)
+            _t['image'] = round((_time.time() - _t_prev) * 1000, 1)
+            _t_prev = _time.time()
 
             # 调试事件：图片选择详情
             if IS_DEV:
@@ -2754,12 +2764,16 @@ def rag():
                 yield f"data: {json.dumps({'type': 'context_built', 'data': {'chunk_count': len(text_used), 'context_length': len(enhanced_context), 'budget_max_chars': CONTEXT_MAX_CHARS, 'min_score_filter': RERANK_CONTEXT_MIN_SCORE, 'confidence_top3': _confidence_score, 'score_stats': {'max': max(_scores) if _scores else 0, 'min': min(_scores) if _scores else 0, 'avg': round(sum(_scores)/len(_scores), 4) if _scores else 0}, 'context_preview': enhanced_context[:500], 'chunks_used': [{'source': ctx.get('meta',{}).get('source',''), 'page': ctx.get('meta',{}).get('page',0), 'score': ctx.get('score',0), 'preview': (ctx.get('doc','') or '')[:100]} for ctx in text_used]}}, ensure_ascii=False)}\n\n"
 
             # 5. 流式生成回答
+            _t['rescue'] = round((_time.time() - _t_prev) * 1000, 1)
+            _t_prev = _time.time()
             from core.engine import get_engine
             engine = get_engine()
 
             for token in engine.generate_answer_stream(message, enhanced_context, history):
                 full_answer.append(token)
                 yield f"data: {json.dumps({'type': 'chunk', 'content': token}, ensure_ascii=False)}\n\n"
+            _t['llm'] = round((_time.time() - _t_prev) * 1000, 1)
+            _t_prev = _time.time()
 
             # 6. P0：答案对齐过滤器
             # 从 LLM 回答中提取图号引用，过滤图片选择结果
@@ -2873,12 +2887,16 @@ def rag():
                     if step.get('name') == 'rerank' and step.get('applied'):
                         rerank_time = step.get('time_ms', 0)
                         rerank_cached = step.get('cached', False)
+                _t['post'] = round((_time.time() - _t_prev) * 1000, 1)
                 finish_event["timing"] = {
                     "total_search_ms": timing_info.get('total_ms', 0),
                     "rerank_ms": rerank_time,
                     "rerank_cached": rerank_cached,
-                    "total_ms": duration_ms
+                    "total_ms": duration_ms,
+                    "stages_ms": _t
                 }
+                # 记录到日志
+                logger.info(f"[性能] 总{duration_ms}ms | 意图{_t['intent']:.0f} 缓存{_t['cache']:.0f} 检索{_t['search']:.0f} VLM{_t['vlm']:.0f} 图片{_t['image']:.0f} 救援{_t['rescue']:.0f} LLM{_t['llm']:.0f} 后处理{_t['post']:.0f}")
             # 10.5 写入语义缓存
             try:
                 from core.semantic_cache import get_semantic_cache

api/sync_routes.py

@@ -315,3 +315,48 @@ def stop_sync_monitor() -> Tuple[Any, int]:
             message="操作失败",
             http_status=500
         )
+
+
+# ==================== 缓存管理 API ====================
+
+@sync_bp.route('/cache/clear', methods=['POST'])
+@require_gateway_auth
+def clear_all_caches() -> Tuple[Any, int]:
+    """
+    清除所有缓存层（Query / Embedding / Rerank / Semantic / Intent 精确缓存）
+
+    用于测试和调试，不影响向量库数据。
+
+    Returns:
+        {"success": true, "message": "缓存已清除", "cleared": [...]}
+    """
+    cleared = []
+
+    # 1. 精确匹配缓存（Query / Embedding / Rerank）
+    try:
+        from core.cache import get_cache_manager
+        cache = get_cache_manager()
+        if cache:
+            cache.clear_all()
+            cleared.append("query_cache")
+            cleared.append("embedding_cache")
+            cleared.append("rerank_cache")
+    except Exception as e:
+        logger.warning(f"清除精确缓存失败: {e}")
+
+    # 2. 语义缓存（FAISS）
+    try:
+        from core.semantic_cache import get_semantic_cache
+        sc = get_semantic_cache()
+        if sc:
+            sc.clear()
+            cleared.append("semantic_cache")
+    except Exception as e:
+        logger.warning(f"清除语义缓存失败: {e}")
+
+    logger.info(f"缓存清除完成: {cleared}")
+    return jsonify({
+        "success": True,
+        "message": f"已清除 {len(cleared)} 层缓存",
+        "cleared": cleared
+    })

core/engine.py

@@ -82,7 +82,7 @@ try:
         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,
+        LLM_TEMPERATURE, LLM_TOP_P, LLM_MAX_TOKENS, RECALL_MULTIPLIER,
         # FAQ 与黑名单
         FAQ_RECALL_TOP_K, FAQ_BOOST_AMOUNT, FAQ_DECAY_MONTHS, FAQ_DECAY_RATE, FAQ_DECAY_MAX,
         BLACKLIST_MIN_DISLIKES, BLACKLIST_CACHE_TTL,
@@ -2361,6 +2361,7 @@ class RAGEngine:
                 "",
                 MODEL,
                 temperature=LLM_TEMPERATURE,
+                top_p=LLM_TOP_P,
                 max_tokens=LLM_MAX_TOKENS,
                 messages=messages,
                 error_prefix="[错误]"

core/intent_analyzer.py

@@ -22,7 +22,7 @@ import threading
 from dataclasses import dataclass, asdict
 from typing import List, Dict, Optional, Tuple
 
-from config import INTENT_TEMPERATURE, INTENT_MAX_TOKENS, INTENT_HISTORY_WINDOW, INTENT_MODEL
+from config import INTENT_TEMPERATURE, INTENT_MAX_TOKENS, INTENT_HISTORY_WINDOW, INTENT_MODEL, INTENT_RESPONSE_FORMAT
 
 logger = logging.getLogger(__name__)
 
@@ -355,17 +355,17 @@ class IntentAnalyzer:
             model = self.model or self._get_default_model()
 
             from core.llm_utils import call_llm
-            content = call_llm(
+            _llm_kwargs = dict(
-                client,
-                "",
-                model,
                 temperature=INTENT_TEMPERATURE,
                 max_tokens=INTENT_MAX_TOKENS,
                 messages=[
                     {"role": "system", "content": self.SYSTEM_PROMPT},
                     {"role": "user", "content": user_prompt}
-                ]
+                ],
             )
+            if INTENT_RESPONSE_FORMAT:
+                _llm_kwargs["response_format"] = {"type": "json_object"}
+            content = call_llm(client, "", model, **_llm_kwargs)
 
             # 解析 JSON
             result = self._parse_json(content)