fix：处理模型错误添加大量文本的bug

app/processor/nodes/input_node.py

@@ -51,5 +51,6 @@ class InputNode(BaseNode):
         updated_state["actions"] = None
         updated_state["clarification"] = None
         updated_state["answer"] = None
+        updated_state["suggestions"] = None
         updated_state["status"] = "running"
         return updated_state

app/processor/nodes/qa_node.py

@@ -18,6 +18,9 @@ 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"
+# 带数据的问答回答完后给的后续请求：意图判断会把"展示一下 9 月销售额"这类说法分到问答，
+# 用户其实想在大屏上看到；回答之后给一个按钮，点了按普通请求走大屏 agent，会话历史里有这次回答
+PUT_ON_SCREEN = "把上面的结果放到大屏上"
 
 
 class QANode(BaseNode):
@@ -93,4 +96,6 @@ class QANode(BaseNode):
 
         updated_state["answer"] = answer
         updated_state["status"] = "completed"
+        if context.data_files:
+            updated_state["suggestions"] = [PUT_ON_SCREEN]
         return updated_state

app/processor/nodes/screen_agent_node.py

@@ -9,6 +9,7 @@ import json
 
 from langchain_core.messages import HumanMessage
 
+from app.actions.batch import mark_sample_data
 from app.agent.screen_agent import (
     AgentReply,
     finalize_reply,
@@ -21,13 +22,19 @@ from app.processor.history import history_messages
 from app.processor.state import QueryState, run_data
 from app.processor.summary import request_summary
 from app.resources.component import get_catalog_cache
+from app.schemas.component_resource import ComponentCatalog
 
 # 评测按这句话识别"草稿因最后一次提交失败没有交付"，单独核对草稿本身是否满足要求
 DRAFT_REJECTED = "修改没有通过校验，这次没有生效，可以换个说法再试"
 
 
-def reply_update(result: dict, *, limit_hit: bool = False) -> QueryState:
-    """把 agent 的运行结果转换成 QueryState 的更新"""
+def reply_update(
+    result: dict, *, limit_hit: bool = False, catalog: ComponentCatalog | None = None
+) -> QueryState:
+    """把 agent 的运行结果转换成 QueryState 的更新
+
+    交付的新增组件按交付后的大屏定 sample_data：留着物料示例数据的也标为示例数据（mark_sample_data）
+    """
     reply: AgentReply | None = result.get("structured_response")
     if reply is None:
         # 收尾也没有给出合法的结论；按真实原因说明，不再一律说超出上限
@@ -44,11 +51,10 @@ def reply_update(result: dict, *, limit_hit: bool = False) -> QueryState:
     if last_write_failed(result.get("messages", [])):
         return {"status": "failed", "answer": DRAFT_REJECTED}
     # 各次成功提交的动作按顺序拼接；clarify、unable 不交付，下次请求从原始大屏重新执行
-    return {
-        "status": "completed",
-        "actions": result.get("actions") or [],
-        "answer": reply.message,
-    }
+    actions = result.get("actions") or []
+    if actions:
+        actions = mark_sample_data(actions, result["working_screen"], catalog)
+    return {"status": "completed", "actions": actions, "answer": reply.message}
 
 
 class ScreenAgentNode(BaseNode):
@@ -91,5 +97,5 @@ class ScreenAgentNode(BaseNode):
             )
             reply = await finalize_reply(result, limit_hit=limit_hit)
             result = {**result, "structured_response": reply}
-        updated_state.update(reply_update(result, limit_hit=limit_hit))
+        updated_state.update(reply_update(result, limit_hit=limit_hit, catalog=catalog))
         return updated_state

app/processor/state.py

@@ -28,6 +28,7 @@ class QueryState(TypedDict):
         ClarificationReply | None
     ]  # 待合并的用户回答或取消指令
     answer: NotRequired[str | None]  # 问答回答，或 agent 给用户的说明
+    suggestions: NotRequired[list[str] | None]  # 可以一键发送的后续请求，前端显示成按钮
     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

app/processor/summary.py

@@ -3,6 +3,8 @@
 只描述本次请求；会话历史由各使用方另行传入，见 history.py
 """
 
+from app.resources.component import get_catalog_cache
+from app.resources.material_fields import identifying_texts
 from app.schemas.base import JsonObject
 from app.schemas.data_file import RequestDataFile
 from app.schemas.input_context import InputContext
@@ -14,25 +16,16 @@ SUMMARY_COLUMNS = 30
 def request_summary(context: InputContext) -> JsonObject:
     """用户输入、澄清记录、编辑上下文、附件、数据文件和大屏摘要"""
     nodes: list[JsonObject] = []
+    catalog = get_catalog_cache()
     for node in context.screen.nodes:
-        # 画面上的标题帮助判断用户指的是哪个组件：图表在 option.title，指标卡在 props.title；
-        # 格式不符时按无标题处理
-        option = node.props.get("option")
-        title = option.get("title") if isinstance(option, dict) else None
-        titles = title if isinstance(title, list) else [title]
-        texts = [
-            item["text"]
-            for item in titles
-            if isinstance(item, dict) and isinstance(item.get("text"), str)
-        ]
-        if isinstance(node.props.get("title"), str) and node.props["title"]:
-            texts.append(node.props["title"])
         nodes.append(
             {
                 "id": node.id,
                 "type": node.type,
                 "name": node.name,
-                "titles": texts,
+                # 画面上的名字或内容帮助判断用户指的是哪个组件；取哪个字段由字段定义的 identifies 决定，
+                # 新物料不用在这里加判断。物料目录没载入时为空
+                "titles": identifying_texts(catalog, node.model_dump(mode="json")),
                 "layout": node.layout.model_dump(mode="json", exclude_unset=True),
             }
         )