feat：完成数据文件上传解析及测试

app/processor/nodes/qa_node.py

@@ -1,14 +1,24 @@
-"""基于当前大屏和会话历史流式回答问题"""
+"""基于当前大屏、上传的数据和会话历史流式回答问题
 
-from langchain_core.messages import HumanMessage, SystemMessage
+由问答 agent 执行：会话里有数据文件时先查询再回答，否则只调用一次模型；只把模型输出的文字推送给前端
+"""
+
+import json
+
+from langchain.agents.middleware.model_call_limit import ModelCallLimitExceededError
+from langchain_core.messages import AIMessageChunk, HumanMessage
 from langchain_core.output_parsers import StrOutputParser
 from langgraph.config import get_stream_writer
 
-from app.clients.llm import get_llm
+from app.agent.data_tools import DataContext
+from app.agent.qa_agent import QAContext, get_qa_agent
 from app.processor.base import BaseNode
 from app.processor.history import history_messages
 from app.processor.state import QueryState
-from app.prompt.qa import QA_SYSTEM_PROMPT
+from app.processor.summary import data_file_summary
+
+# 模型在一次回答里分几次输出文字（例如查询前后各说一段）时，段落之间空一行
+PARAGRAPH_BREAK = "\n\n"
 
 
 class QANode(BaseNode):
@@ -23,23 +33,42 @@ class QANode(BaseNode):
         except RuntimeError:
             writer = None
 
+        context = state["input_context"]
+        payload = context.model_dump(
+            mode="json", exclude_unset=True, exclude={"history", "data_files"}
+        )
+        if context.data_files:
+            # 只给表名、行数和列名，文件路径和列概况不进提示词，需要时由 describe_data 查看
+            payload["data_files"] = [
+                data_file_summary(item) for item in context.data_files
+            ]
+        agent = get_qa_agent(with_data=bool(context.data_files))
+
         parser = StrOutputParser()
         parts: list[str] = []
-        context = state["input_context"]
-        async for chunk in get_llm().astream(
-            [
-                SystemMessage(content=QA_SYSTEM_PROMPT),
-                # 会话历史作为此前的对话消息，本次请求是最后一条用户消息
-                *history_messages(context.history),
-                HumanMessage(
-                    content=context.model_dump_json(
-                        exclude_unset=True, exclude={"history"}
-                    )
-                ),
-            ]
-        ):
-            # 工具调用等无文本片段会得到空串
-            if delta := parser.invoke(chunk):
+        last_message_id = None
+        try:
+            async for chunk, metadata in agent.astream(
+                {
+                    "messages": [
+                        # 会话历史作为此前的对话消息，本次请求是最后一条用户消息
+                        *history_messages(context.history),
+                        HumanMessage(content=json.dumps(payload, ensure_ascii=False)),
+                    ]
+                },
+                context=QAContext(data=DataContext(tuple(context.data_files))),
+                stream_mode="messages",
+            ):
+                # 工具结果、工具调用和中间件插入的消息都不推送
+                if metadata.get("langgraph_node") != "model" or not isinstance(
+                    chunk, AIMessageChunk
+                ):
+                    continue
+                if not (delta := parser.invoke(chunk)):
+                    continue
+                if parts and chunk.id != last_message_id:
+                    delta = PARAGRAPH_BREAK + delta
+                last_message_id = chunk.id
                 parts.append(delta)
                 if writer is not None:
                     writer(
@@ -49,6 +78,15 @@ class QANode(BaseNode):
                             "delta": delta,
                         }
                     )
+        except ModelCallLimitExceededError:
+            self.logger.warning("问答 agent 超出模型调用上限（MAX_QA_STEPS）")
+            updated_state["status"] = "failed"
+            updated_state["answer"] = (
+                "查询步骤超出上限，这次没有完成，可以缩小范围或换个问法再试"
+            )
+            return updated_state
+
+        # 推送了什么就保存什么，会话历史和前端看到的回答一致
         answer = "".join(parts).strip()
         # 没有有效文本时报错，避免将空回答标记为处理完成
         if not answer: