update

app/chat_qa_query/api/router/query_router.py

@@ -7,11 +7,13 @@ from watchfiles import awatch
 from app.chat_qa_query.api.qa_dependencies import get_qa_query_service
 from app.chat_qa_query.api.schema.query_qa_schema import QueryQASchema
 from app.chat_qa_query.service.query_qa_service import QueryQAService
+from app.core.log import logger
 
 query_qa_router = APIRouter()
 
 
 @query_qa_router.post("/api/qa/query")
 async def query_handler(query: QueryQASchema, query_service: Annotated[QueryQAService, Depends(get_qa_query_service)]):
+    logger.info(f"请求参数:{query.external_str},历史top15:{query.messages_top15}")
     return await query_service.query(query.external_str, query.messages_top15)
 

app/chat_qa_query/query_graph.py

@@ -45,7 +45,7 @@ if __name__ == '__main__':
             archive_messages_mysql_repository = ArchiveMessagesRepository(db_session)
             qa_milvus_repository = QARepository()
 
-            state: QueryQAState = QueryQAState(original_query="培训的费用？")
+            state: QueryQAState = QueryQAState(original_query="暂时有点忙", history=['2026-05-26T13:14:59 EXTERNAL 暂时有点忙\n', '2026-05-26T12:48:09 INTERNAL 下午好呀 你这边怎么计划的呀 考虑跟咱们6月1号的班级学习嘛\n'])
             context = QueryQAContext(meta_mysql_repository=archive_messages_mysql_repository,
                                 qa_milvus_repository=qa_milvus_repository
                                 )

app/chat_qa_query/query_nodes/node_rewrite.py

@@ -28,7 +28,6 @@ problem_rewriting_prompt = """
 
 4. **输出格式**：
    - 仅输出JSON，不要任何解释性文字
-   - rewritten_query 必须是一个完整的疑问句（以"是什么"、"有哪些"、"怎么样"、"如何"等结尾，或以问号结尾）
 
 ## 改写示例
 
@@ -76,6 +75,9 @@ problem_rewriting_prompt = """
 历史会话：
 {history_text}
 
+历史会话字段说明：
+EXTERNAL：客户
+INTERNAL：销售
 当前问题：{query}
 
 请直接输出JSON：
@@ -90,6 +92,7 @@ async def rewrite_query(state:QueryQAState):
     """
     history_text = state.get("history", "")
     query = state.get("original_query", "")
+    logger.info(f"问题改写：{query},历史会话：{history_text}")
     prompt = PromptTemplate(template=problem_rewriting_prompt, input_variables=["history_text", "query"])
     output = JsonOutputParser()
     chain = prompt | llm | output

app/chat_qa_query/query_nodes/node_search_embedding.py

@@ -15,7 +15,7 @@ async def search_embed(state:QueryQAState) :
    batch_embeddings = await embedding_client.aembed_query(rewritten_query)
    qa_repository = QARepository()
 
-   result = await qa_repository.search_qa(query_text=batch_embeddings, top_k=3)
+   result = await qa_repository.search_qa(query_vector=batch_embeddings, top_k=3)
    logger.info(f"原始问题:{original_query},改写后的问题：{rewritten_query},匹配结果：{result}")
    if result:
        question = result[0]["question"]

app/repository/milvus/message_qa_repository.py

@@ -100,12 +100,12 @@ class QARepository:
             logger.error(f"QA数据入库失败: {str(e)}", exc_info=True)
             return []
 
-    async def search_qa(self, query_text: List[float], top_k: int = 5) -> List[Dict[str, Any]]:
+    async def search_qa(self, query_vector: List[float], top_k: int = 5) -> List[Dict[str, Any]]:
         """
         根据查询文本搜索QA数据
 
         Args:
-            query_text: 查询文本
+            query_vector: 查询向量
             top_k: 返回结果数量
 
         Returns:
@@ -114,13 +114,16 @@ class QARepository:
         res = milvus_client.client.search(
             collection_name=self.collection_name,
             anns_field="question_dense_vector",
-            data=[query_text],
+            data=[query_vector],
             limit=top_k,
-            search_params={"metric_type": "COSINE"},
+            search_params={
+                "metric_type": "COSINE",
+                "efSearch": 300  # 你配置 efSearch=200：检索时最多遍历 200 个候选节点，再从中选出相似度最高的结果。
+            },
             output_fields=["question", "answer", "created_at"]
         )
 
-        score_threshold = 0.6
+        score_threshold = 0.60
         filter_results = []
 
         # res 是二维结构：[ [单条结果列表] ]

readme

@@ -0,0 +1,5 @@
+  启动：nohup uvicorn app.uvicorn_main:app --host 0.0.0.0 --port 8000 --workers 4 > uvicorn.log 2>&1 &
+  打包：py package.py
+  同步库+虚拟环境：uv sync
+
+