Implement follow-up detection in dialog context to prevent context pollution in multi-topic conversations. Introduce is_likely_follow_up function to determine when to inject session history based on user input. Adjust _load_session_text2sql_context to conditionally include context based on follow-up status, enhancing SQL generation accuracy. Update relevant API endpoints to pass user text for improved context handling.

backend/utils/dialog_context.py

@@ -6,6 +6,7 @@ from __future__ import annotations
 
 import json
 import logging
+import re
 from typing import Any, Dict, List, Tuple
 
 logger = logging.getLogger(__name__)
@@ -14,6 +15,28 @@ _MAX_ASSISTANT_SQL_CHARS = 1400
 _MAX_ASSISTANT_TEXT_CHARS = 900
 _MAX_BLOCK_CHARS = 7500
 
+# 续问/指代常见触发词：用于判定是否需要注入会话上文，避免多话题串扰。
+# 规则应偏保守：宁可把新话题当作“无上文”也不要带入无关上下文污染 SQL。
+_FOLLOW_UP_PATTERN = re.compile(
+    r"(再|继续|同样|沿用|还是|照旧|刚才|上面|上一条|上一个|上次|上述|这个SQL|这条SQL|按上面|按刚才|在此基础上|同一口径|同口径|按之前|改成|改为|加上|加一下|筛一下|过滤一下|补充|补一下|追加|优化一下|调整一下|换成|换为)",
+    re.IGNORECASE,
+)
+
+
+def is_likely_follow_up(user_text: str) -> bool:
+    """
+    仅基于用户本轮文本做“是否续问”的轻量判断。
+    True  -> 允许注入少量上文（最近 1~2 轮），用于指代消解/沿用口径
+    False -> 新话题，禁用上文，避免上下文污染
+    """
+    t = (user_text or "").strip()
+    if not t:
+        return False
+    # 很短的“再来一个/同上/继续”等通常是续问
+    if len(t) <= 10 and _FOLLOW_UP_PATTERN.search(t):
+        return True
+    return bool(_FOLLOW_UP_PATTERN.search(t))
+
 
 def _truncate(s: str, max_len: int) -> str:
     s = (s or "").strip()