fix: skip semantic cache for short follow-up queries to avoid false hits

Short follow-ups like "B3类呢？" are too context-dependent for embedding
similarity matching. Only exact-match cache is used when history exists
and query is under 20 chars.

core/intent_analyzer.py

@@ -309,7 +309,9 @@ class IntentAnalyzer:
 
         # 2. 尝试从语义缓存获取
         cache = self._get_cache()
-        if 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
+        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)
@@ -325,7 +327,10 @@ class IntentAnalyzer:
             else:
                 logger.warning("意图分析缓存: _get_embedding 返回 None，跳过缓存")
         else:
-            logger.warning("意图分析缓存: _get_cache 返回 None，缓存不可用")
+            if is_short_followup:
+                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