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）

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(
-                client,
-                "",
-                model,
+            _llm_kwargs = dict(
                 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)