feat：LLM 分析文件并生成结构化建议

app/services/agent.py

@@ -7,13 +7,16 @@ from uuid import uuid4
 from langgraph.types import Command
 from loguru import logger
 
+from app.agent.data_tools import DataContext
 from app.config import settings
 from app.core.exceptions import BizException
 from app.datafiles.service import DataFileService
 from app.observability.tracing import astream_screen
 from app.processor.history import trim_history
+from app.processor.state import RunContext
 from app.schemas.agent import AgentResult, ResumeRequest
 from app.schemas.conversation import Turn
+from app.schemas.data_file import DataFileRecord, RequestDataFile
 from app.schemas.request import DesignRequest
 from app.services.conversation import ConversationStore
 
@@ -43,8 +46,9 @@ class AgentService:
                 conversation_id, settings.conversation_history_turns
             )
             # 附件引用的文件不属于本会话时在流开始前返回 404
-            data_files = await self.data_files.request_files(
-                user_id, conversation_id, request.attachments
+            attached = request.data_file_ids
+            records = await self.data_files.request_files(
+                user_id, conversation_id, attached
             )
         except BaseException:
             self.release(thread_id)
@@ -64,7 +68,16 @@ class AgentService:
             settings.history_max_tokens,
         )
         # 同样只信任服务端读取的结果；会话内的文件每一轮都可用，不附带也能查询
-        payload["data_files"] = [item.model_dump(mode="json") for item in data_files]
+        # 请求会进检查点，只放摘要；工具要用的列概况和路径通过运行上下文传给图
+        payload["data_files"] = [
+            RequestDataFile(
+                file_id=record.file_id,
+                name=record.name,
+                attached_this_turn=record.file_id in attached,
+                tables=record.tables,
+            ).model_dump(mode="json")
+            for record in records
+        ]
 
         return (
             thread_id,
@@ -76,6 +89,7 @@ class AgentService:
             },
             conversation_id,
             user_id,
+            self._run_context(records),
         )
 
     async def prepare_resume(self, request: ResumeRequest, user_id: int):
@@ -117,6 +131,10 @@ class AgentService:
                         )
                     }
                 )
+            # 运行上下文不在检查点里，恢复时按会话重新读取；期间删除或过期的文件工具查不到
+            records = await self.data_files.request_files(
+                user_id, conversation_id, set()
+            )
 
             # 每次有效补充或取消生成新的输入 ID，任务和会话保持不变
             return (
@@ -127,11 +145,15 @@ class AgentService:
                 ),
                 conversation_id,
                 user_id,
+                self._run_context(records),
             )
         except BaseException:
             self.release(thread_id)
             raise
 
+    def _run_context(self, records: list[DataFileRecord]) -> RunContext:
+        return RunContext(data=DataContext(tuple(records), self.data_files.root))
+
     def release(self, thread_id):
         self._active.discard(thread_id)
         self._conversation_of.pop(thread_id, None)
@@ -199,7 +221,7 @@ class AgentService:
             clarification=state.get("clarification") if waiting else None,
         )
 
-    async def stream(self, thread_id, payload, session_id, user_id=None):
+    async def stream(self, thread_id, payload, session_id, user_id=None, context=None):
         """消费图事件，完整状态只用于整理最终结果，不直接暴露给客户端"""
         state = None
         with logger.contextualize(thread_id=thread_id):
@@ -210,6 +232,7 @@ class AgentService:
                     thread_id=thread_id,
                     session_id=session_id,
                     user_id=user_id,
+                    context=context,
                 )
             ) as events:
                 async for mode, data in events: