diff --git a/gfa_core/extension_stage_lightweight/nl_intent.py b/gfa_core/extension_stage_lightweight/nl_intent.py
index 5026f1562..964b237f5 100644
--- a/gfa_core/extension_stage_lightweight/nl_intent.py
+++ b/gfa_core/extension_stage_lightweight/nl_intent.py
@@ -25,6 +25,7 @@ MAX_JOB_FILE_HINTS = 200
 
 _ENGLISH_STOPWORDS = {"analysis", "data", "file", "run", "input", "output", "files", "using", "tool"}
 _CHINESE_STOPWORDS = {"分析", "序列", "文件", "数据", "蛋白", "拓展", "扩展", "工具", "进行", "搜索"}
+_STRONG_TRIGGER_TOKENS = ("拓展分析", "扩展分析", "extension")
 
 _RESPONSE_SYSTEM_PROMPT = (
     "你是植物基因组分析平台的自然语言意图解析器。"
@@ -139,22 +140,28 @@ def _record_keyword_sources(record: Mapping[str, Any]) -> list[str]:
 def _rough_candidates(message: str, records: list[dict]) -> list[str]:
     """按关键词粗筛可能命中的记录 id；不含 LLM 调用。
 
-    返回按命中关键词数降序排列（同分按 id 排序）；「拓展分析/扩展分析/extension」
-    这类不含功能线索的强触发词才允许全部候选进目录，其余消息只返回命中子集。
+    union 策略：先剔除强触发词（拓展分析/扩展分析/extension）后按关键词源
+    匹配子集，命中数降序（同分按 id 排序）；除强触发词外没有任何关键词命中
+    （纯「拓展分析」）时才返回全量候选兜底，否则返回 []。
     """
     text = str(message or "").lower()
-    if any(token in text for token in ("拓展分析", "扩展分析", "extension")):
-        return [str(record.get("extension_id")) for record in records if record.get("extension_id")]
+    stripped = text
+    for token in _STRONG_TRIGGER_TOKENS:
+        stripped = stripped.replace(token, " ")
     scored: list[tuple[int, str]] = []
     for record in records:
         extension_id = str(record.get("extension_id") or "").strip()
         if not extension_id:
             continue
-        hits = sum(1 for keyword in _record_keyword_sources(record) if keyword and keyword in text)
+        hits = sum(1 for keyword in _record_keyword_sources(record) if keyword and keyword in stripped)
         if hits:
             scored.append((hits, extension_id))
-    scored.sort(key=lambda item: (-item[0], item[1]))
-    return [extension_id for _, extension_id in scored]
+    if scored:
+        scored.sort(key=lambda item: (-item[0], item[1]))
+        return [extension_id for _, extension_id in scored]
+    if any(token in text for token in _STRONG_TRIGGER_TOKENS):
+        return [str(record.get("extension_id")) for record in records if record.get("extension_id")]
+    return []
 
 
 def compact_extension_name(extension_id: str) -> str: