feat：LLM 分析文件并生成结构化建议

app/processor/graph.py

@@ -11,7 +11,7 @@ from app.processor.nodes.intent_node import IntentNode
 from app.processor.nodes.qa_node import QANode
 from app.processor.nodes.screen_agent_node import ScreenAgentNode
 from app.processor.nodes.wait_for_user_node import WaitForUserNode
-from app.processor.state import QueryState
+from app.processor.state import QueryState, RunContext
 
 
 def build_checkpoint_serde() -> JsonPlusSerializer:
@@ -47,7 +47,7 @@ def route_after_wait(state: QueryState) -> str:
 
 
 def build_graph(checkpointer=None):
-    workflow = StateGraph(QueryState)
+    workflow = StateGraph(QueryState, context_schema=RunContext)
     nodes = {
         "input_node": InputNode(),
         "intent_node": IntentNode(),

app/processor/nodes/qa_node.py

@@ -10,11 +10,10 @@ from langchain_core.messages import AIMessageChunk, HumanMessage
 from langchain_core.output_parsers import StrOutputParser
 from langgraph.config import get_stream_writer
 
-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.processor.state import QueryState, run_data
 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
@@ -56,7 +55,7 @@ class QANode(BaseNode):
                         HumanMessage(content=json.dumps(payload, ensure_ascii=False)),
                     ]
                 },
-                context=QAContext(data=DataContext(tuple(context.data_files))),
+                context=QAContext(data=run_data()),
                 stream_mode="messages",
             ):
                 # 工具结果、工具调用和中间件插入的消息都不推送

app/processor/nodes/screen_agent_node.py

@@ -7,12 +7,11 @@ import json
 
 from langchain_core.messages import HumanMessage
 
-from app.agent.data_tools import DataContext
 from app.agent.screen_agent import AgentReply, get_screen_agent
 from app.agent.tools import ScreenAgentContext
 from app.processor.base import BaseNode
 from app.processor.history import history_messages
-from app.processor.state import QueryState
+from app.processor.state import QueryState, run_data
 from app.processor.summary import request_summary
 from app.resources.component import get_catalog_cache
 
@@ -68,7 +67,7 @@ class ScreenAgentNode(BaseNode):
             context=ScreenAgentContext(
                 screen=context.screen,
                 catalog=catalog,
-                data=DataContext(tuple(context.data_files)),
+                data=run_data(),
             ),
         )
         if result.get("structured_response") is None:

app/processor/state.py

@@ -1,5 +1,9 @@
+from dataclasses import dataclass, field
 from typing import Literal, NotRequired, TypedDict
 
+from langgraph.runtime import get_runtime
+
+from app.agent.data_tools import DataContext
 from app.schemas.actions import Action
 from app.schemas.base import JsonObject
 from app.schemas.clarification import Clarification
@@ -25,3 +29,21 @@ class QueryState(TypedDict):
     status: NotRequired[
         Literal["running", "waiting_for_input", "completed", "cancelled", "failed"]
     ]  # 后端处理、等待补充、处理完成、用户取消或失败，完成不代表前端已执行
+
+
+@dataclass(frozen=True)
+class RunContext:
+    """图的运行上下文：本次运行要用、但不进检查点的依赖，每次启动和恢复时由调用方传入"""
+
+    # 本会话的数据文件；状态里的 input_context.data_files 只有摘要，列概况和路径在这里
+    data: DataContext = field(default_factory=DataContext)
+
+
+def run_data() -> DataContext:
+    """当前运行可用的数据表；单独调用节点（不在图里执行）或调用方没传 context 时为空"""
+    try:
+        runtime = get_runtime(RunContext)
+    except RuntimeError:
+        return DataContext()
+    context = runtime.context if runtime is not None else None
+    return context.data if context is not None else DataContext()