fix(chat): 拦截无依据的非业务问题

review_agent/services.py

@@ -108,10 +108,13 @@ def send_message(conversation: Conversation, content: str) -> tuple[Message, Mes
 
     user_message = append_user_message(conversation, content)
     knowledge_context = build_knowledge_context(content)
-    try:
-        reply_content = generate_reply(conversation, content, knowledge_context=knowledge_context)
-    except (LLMConfigurationError, LLMRequestError) as exc:
-        reply_content = f"模型调用失败：{exc}"
+    if should_refuse_ungrounded_chat(conversation, content, knowledge_context):
+        reply_content = out_of_scope_reply()
+    else:
+        try:
+            reply_content = generate_reply(conversation, content, knowledge_context=knowledge_context)
+        except (LLMConfigurationError, LLMRequestError) as exc:
+            reply_content = f"模型调用失败：{exc}"
 
     assistant_message = append_assistant_message(conversation, reply_content)
 
@@ -127,6 +130,31 @@ def stream_message(conversation: Conversation, content: str):
 
     user_message = append_user_message(conversation, content)
     assistant_parts: list[str] = []
+    knowledge_context = build_knowledge_context(content)
+
+    if should_refuse_ungrounded_chat(conversation, content, knowledge_context):
+        reply_content = out_of_scope_reply()
+        assistant_message = append_assistant_message(conversation, reply_content)
+        yield sse_event(
+            "meta",
+            {
+                "conversation_id": conversation.pk,
+                "title": conversation.title or build_conversation_title(content),
+                "user_message_id": user_message.pk,
+                "user_message": user_message.content,
+            },
+        )
+        yield sse_event("chunk", {"delta": reply_content})
+        yield sse_event(
+            "done",
+            {
+                "assistant_message_id": assistant_message.pk,
+                "conversation_id": conversation.pk,
+                "title": conversation.title,
+            },
+        )
+        return
+
     route = route_message_intent(conversation, content)
     logger.info(
         "Stream message started",
@@ -395,7 +423,6 @@ def stream_message(conversation: Conversation, content: str):
 
     stream_failed = False
     stream_error = ""
-    knowledge_context = build_knowledge_context(content)
     try:
         for chunk in stream_reply(conversation, content, knowledge_context=knowledge_context):
             assistant_parts.append(chunk)
@@ -497,6 +524,76 @@ def build_knowledge_context(content: str, *, n_results: int = 5) -> str:
     return "\n\n".join(lines)
 
 
+def should_refuse_ungrounded_chat(
+    conversation: Conversation,
+    content: str,
+    knowledge_context: str = "",
+) -> bool:
+    if (knowledge_context or "").strip():
+        return False
+    if _is_business_related_question(content):
+        return False
+    if _has_active_attachments(conversation):
+        return False
+    return True
+
+
+def out_of_scope_reply() -> str:
+    return (
+        "没有在当前启用的知识库材料中找到可依据的内容，且这个问题与当前主营业务无关。"
+        "为避免编造，我不能直接回答。请先上传或启用相关知识库材料，或改问体外诊断试剂注册资料审核、"
+        "文件汇总、法规核查、申报填表等业务范围内的问题。"
+    )
+
+
+def _is_business_related_question(content: str) -> bool:
+    normalized = (content or "").lower()
+    compact = "".join(normalized.split())
+    if not compact:
+        return True
+    business_keywords = [
+        "审核智能体",
+        "体外诊断",
+        "ivd",
+        "nmpa",
+        "cmde",
+        "医疗器械",
+        "注册资料",
+        "注册申报",
+        "注册检验",
+        "注册证",
+        "申报资料",
+        "申报文件",
+        "法规",
+        "核查",
+        "审评",
+        "审核",
+        "整改",
+        "风险",
+        "说明书",
+        "临床",
+        "性能",
+        "安全",
+        "适用范围",
+        "预期用途",
+        "附件",
+        "文件",
+        "压缩包",
+        "目录",
+        "页数",
+        "清单",
+        "汇总",
+        "模板",
+        "填表",
+        "知识库",
+        "检索",
+        "报告",
+        "材料",
+        "资料",
+    ]
+    return any(keyword in compact for keyword in business_keywords)
+
+
 def build_filename_matched_document_context(query: str, *, max_chars: int = 12000) -> str:
     terms = _knowledge_query_terms(query)
     if not terms: