fix(chat): 拦截无依据的非业务问题

2026-06-09 08:23:08 +08:00
parent 42187bf8e9
commit 26e675e5d3
3 changed files with 168 additions and 7 deletions
--- a/review_agent/services.py
+++ b/review_agent/services.py
@@ -108,6 +108,9 @@ def send_message(conversation: Conversation, content: str) -> tuple[Message, Mes

    user_message = append_user_message(conversation, content)
    knowledge_context = build_knowledge_context(content)
+    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:
@@ -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:
--- a/tests/test_chat_knowledge_context.py
+++ b/tests/test_chat_knowledge_context.py
@@ -1,7 +1,7 @@
 import pytest

 from review_agent.models import KnowledgeBaseDocument
-from review_agent.services import build_knowledge_context
+from review_agent.services import build_knowledge_context, send_message, stream_message


 pytestmark = pytest.mark.django_db
@@ -57,3 +57,67 @@ def test_build_knowledge_context_uses_full_document_when_name_matches(settings,
    assert "全文材料" in context
    assert "来源：用户知识库/孙之烨-260510.txt" in context
    assert "完整经历：曾组织技术分享并带队参加竞赛" in context
+
+
+def test_send_message_refuses_out_of_scope_answer_without_knowledge_context(monkeypatch, django_user_model):
+    from review_agent.models import Conversation
+
+    user = django_user_model.objects.create_user(username="owner", password="pass")
+    conversation = Conversation.objects.create(user=user, title="会话")
+    monkeypatch.setattr(
+        "review_agent.services.search_knowledge_base",
+        lambda query, n_results=5: {"query": query, "results": [], "error_message": ""},
+    )
+    monkeypatch.setattr(
+        "review_agent.services.generate_reply",
+        lambda *args, **kwargs: pytest.fail("out-of-scope answer without knowledge context must not call LLM"),
+    )
+
+    _, assistant_message = send_message(conversation, "孙之烨是谁")
+
+    assert "没有在当前启用的知识库材料中找到" in assistant_message.content
+    assert "与当前主营业务无关" in assistant_message.content
+
+
+def test_stream_message_refuses_out_of_scope_answer_without_knowledge_context(monkeypatch, django_user_model):
+    from review_agent.models import Conversation
+
+    user = django_user_model.objects.create_user(username="owner", password="pass")
+    conversation = Conversation.objects.create(user=user, title="会话")
+    monkeypatch.setattr(
+        "review_agent.services.search_knowledge_base",
+        lambda query, n_results=5: {"query": query, "results": [], "error_message": ""},
+    )
+    monkeypatch.setattr(
+        "review_agent.services.stream_reply",
+        lambda *args, **kwargs: pytest.fail("out-of-scope answer without knowledge context must not call streaming LLM"),
+    )
+    monkeypatch.setattr(
+        "review_agent.services.generate_reply",
+        lambda *args, **kwargs: pytest.fail("out-of-scope answer without knowledge context must not call fallback LLM"),
+    )
+
+    frames = list(stream_message(conversation, "给我一份红烧肉菜谱"))
+
+    assert any("没有在当前启用的知识库材料中找到" in frame for frame in frames)
+    assert any("与当前主营业务无关" in frame for frame in frames)
+    assert any("done" in frame for frame in frames)
+
+
+def test_business_question_without_knowledge_context_can_use_llm(monkeypatch, django_user_model):
+    from review_agent.models import Conversation
+
+    user = django_user_model.objects.create_user(username="owner", password="pass")
+    conversation = Conversation.objects.create(user=user, title="会话")
+    monkeypatch.setattr(
+        "review_agent.services.search_knowledge_base",
+        lambda query, n_results=5: {"query": query, "results": [], "error_message": ""},
+    )
+    monkeypatch.setattr(
+        "review_agent.services.generate_reply",
+        lambda *args, **kwargs: "注册检验报告通常用于证明产品性能符合要求。",
+    )
+
+    _, assistant_message = send_message(conversation, "注册检验报告有什么作用")
+
+    assert "注册检验报告" in assistant_message.content
--- a/tests/test_file_summary_workflow.py
+++ b/tests/test_file_summary_workflow.py
@@ -286,7 +286,7 @@ def test_stream_message_falls_back_to_non_stream_reply_when_stream_breaks(monkey
        lambda conversation, content, knowledge_context="": "非流式完整回复",
    )

-    frames = list(stream_message(conversation, "普通问题"))
+    frames = list(stream_message(conversation, "注册检验报告审核要点有哪些"))

    joined = "".join(frames)
    assert "已生成部分内容" in joined