模型反馈修正

2025-11-07 15:30:27 +08:00
parent 3f7cd23bdc
commit 6fe29a51ce
6 changed files with 166 additions and 34 deletions
--- a/graph_chat/gen_data_report_agent.py
+++ b/graph_chat/gen_data_report_agent.py
@@ -1,12 +1,17 @@
 import json
 import logging
+from datetime import datetime
 from typing import TypedDict, Optional

+import orjson
+from anyio import current_time
 from langgraph.checkpoint.memory import MemorySaver
 from langgraph.graph import StateGraph, END, START

 from graph_chat.gen_sql_chart_agent import gen_history_llm
+from util.utils import extract_nested_json

+from service.conversation_service import update_conversation
 logger = logging.getLogger(__name__)
 from template.template import get_base_template

@@ -20,41 +25,103 @@ class DateReportAgentState(TypedDict):
    run_sql_error: str
    retry_count: int
    question:str
+    retry_sql:str
+    sql_correct:bool
+


 def _run_sql(state: DateReportAgentState) -> dict:
    from main_service import vn
    sql = state['sql']
+    retry_sql = state.get('retry_sql', '')
+    sql_correct = state.get('sql_correct', False)
+    logger.info(f"user:{state.get('user_id', '1')} 进入 _run_sql 节点")
    try:
+        if retry_sql and not sql_correct:
+            sql = retry_sql
+        update_conversation(id=state.get('id'), sql=sql)
        df = vn.run_sql_2(sql)
        result = df.to_dict(orient='records')
+        logger.info(f"run sql result {result}")
        retry = state.get("retry_count", 0) + 1
-        return {'data': result, 'retry_count': retry}
+        logger.info(f"old_sql:{sql} retry sql {retry_sql}")
+        return {'data': result, 'retry_count': retry, 'retry_sql': retry_sql}
    except Exception as e:
        retry = state.get("retry_count", 0) + 1
        logger.error(f"Error running sql: {sql}, error: {e}")
        return {'retry_count': retry, 'run_sql_error': '运行sql语句失败，请检查语法或联系管理员。'}


+def _feedback_qa(state: DateReportAgentState) -> dict:
+    logger.info(f"user:{state.get('user_id', '1')} 进入 _feedback_qa 节点")
+    try:
+        user_question = state['question']
+        sql = state['sql']
+        sql_result = state['data']
+        template = get_base_template()
+        feedback_temp = template['template']['result_feedback']
+        logger.info(f"feedback_temp is {feedback_temp}")
+        sys_promot = feedback_temp['system'].format(question=user_question, sql=sql, sql_result=sql_result,
+                                                    current_time=datetime.now())
+        logger.info(f"system_temp is {sys_promot}")
+        result = gen_history_llm.invoke(sys_promot).text()
+        logger.info(f"feedback result: {result}")
+        result = extract_nested_json(result)
+        result = orjson.loads(result)
+        logger.info(f"提取json成功")
+        logger.info(f"result is {result} type:{type(result)}")
+        print("result is {0}".format(result.keys()))
+        if not result["is_result_correct"] and (result["suggested_sql"]):
+            logger.info("开始替换")
+            logger.info(f"suggested_sql is ".format(result["suggested_sql"]))
+            # state["sql"] = result["suggested_sql"]
+            # state["retry_sql"] = True
+            logger.info(f"current state: {state}")
+            return {"retry_sql": result["suggested_sql"]}
+        if result["is_result_correct"]:
+            return {"sql_correct": True}
+    except Exception as e:
+        logger.error(f"Error feedback: {e}")
+
+
+
 def run_sql_hande(state: DateReportAgentState) -> str:
-    sql_error = state.get('run_sql_error', '')
+    # sql_error = state.get('run_sql_error', '')
    data = state.get('data', {})
+    sql_correct = state.get('sql_correct', False)
    sql_retry_count = state.get('retry_count', 0)
-    if sql_error and len(sql_error) > 0:
-        if sql_retry_count < 2:
-            return '_run_sql'
+    logger.info("sql_retry_count is {0}".format(sql_retry_count))
+
+    if sql_retry_count < 2:
+        if sql_correct:
+            return '_gen_report'
        else:
-            return END
-    if data and len(data) > 0:
-        return '_gen_report'
+            return '_feedback_qa'
    else:
-        if sql_retry_count < 2:
-            return '_run_sql'
+        if data and len(data) > 0:
+            return '_gen_report'
        else:
-            return END
+            END
+    # retry_sql = state.get('retry_sql', False)
+
+    # if sql_error and len(sql_error) > 0:
+    #     if sql_retry_count < 2:
+    #         return '_feedback_qa'
+    #         # return '_run_sql'
+    #     else:
+    #         return END
+    #
+    # if data and len(data) > 0:
+    #     return '_gen_report'
+    # else:
+    #     if sql_retry_count < 2:
+    #         return '_feedback_qa'
+    #     else:
+    #         return END


 def _gen_report(state: DateReportAgentState) -> dict:
+    logger.info(f"user:{state.get('user_id', '1')} 进入 _gen_report 节点")
    sql = state['sql']
    question = state['question']
    run_sql_error = state.get('run_sql_error', '')
@@ -62,15 +129,20 @@ def _gen_report(state: DateReportAgentState) -> dict:
    template = get_base_template()
    result_summary = template['template']['result_summary']['system']
    pro = result_summary.format(sql=sql, error_message=run_sql_error, question=question, data=data)
-    txt = gen_history_llm.invoke(pro).text
+    txt = gen_history_llm.invoke(pro).text()
    return {'summary': txt}


 workflowx = StateGraph(DateReportAgentState)
 workflowx.add_node("_run_sql", _run_sql)
+workflowx.add_node("_feedback_qa", _feedback_qa)
 workflowx.add_node("_gen_report", _gen_report)
 workflowx.add_edge(START, "_run_sql")
+workflowx.add_edge("_feedback_qa", "_run_sql")
 workflowx.add_edge("_gen_report", END)
-workflowx.add_conditional_edges("_run_sql", run_sql_hande, [END, '_gen_report'])
+workflowx.add_conditional_edges("_run_sql", run_sql_hande, [END, '_gen_report','_feedback_qa'])
 memory = MemorySaver()
 result_report_agent = workflowx.compile(checkpointer=memory)
+png_data=result_report_agent .get_graph().draw_mermaid_png()
+with open("D://graph2.png", "wb") as f:
+    f.write(png_data)
--- a/graph_chat/gen_sql_chart_agent.py
+++ b/graph_chat/gen_sql_chart_agent.py
@@ -65,7 +65,7 @@ def _rewrite_user_question(state: SqlAgentState) -> dict:
    rewrite_temp = template['template']['rewrite_question']
    history = state.get('history', [])
    sys_promot = rewrite_temp['system'].format(current_question=user_question, history=history)
-    new_question = gen_history_llm.invoke(sys_promot).text
+    new_question = gen_history_llm.invoke(sys_promot).text()
    logger.info("new_question:{0}".format(new_question))
    result = extract_nested_json(new_question)
    logger.info(f"result:{result}")
@@ -93,7 +93,7 @@ def _gen_sql(state: SqlAgentState) -> dict:
        user_temp = sql_temp['user'].format(question=question,
                                            current_time=datetime.now().strftime('%Y-%m-%d %H:%M:%S'))
        rr = gen_sql_llm.invoke(
-            [{'role': 'system', 'content': sys_temp}, {'role': 'user', 'content': user_temp}]).text
+            [{'role': 'system', 'content': sys_temp}, {'role': 'user', 'content': user_temp}]).text()
        result = extract_nested_json(rr)
        logger.info(f"gensql result: {result}")
        result = orjson.loads(result)
@@ -120,7 +120,7 @@ def _gen_chart(state: SqlAgentState) -> dict:
                                                   lang='中文', sql=sql, chart_type=char_type)
        user_char_temp = char_temp['user'].format(sql=sql, chart_type=char_type, question=question)
        rr = gen_history_llm.invoke(
-            [{'role': 'system', 'content': sys_char_temp}, {'role': 'user', 'content': user_char_temp}]).text
+            [{'role': 'system', 'content': sys_char_temp}, {'role': 'user', 'content': user_char_temp}]).text()
        retry = state.get("chart_retry_count", 0) + 1
        return {"chart_retry_count": retry, 'gen_chart_result': orjson.loads(extract_nested_json(rr))}
    except Exception as e:
--- a/main_service.py
+++ b/main_service.py
@@ -7,7 +7,7 @@ import util.utils
 from logging_config import LOGGING_CONFIG
 from service.cus_vanna_srevice import CustomVanna, QdrantClient, TTLCacheWrapper
 from service.question_feedback_service import save_save_question_async,query_predefined_question_list
-from service.conversation_service import save_conversation,update_conversation,get_sql_by_id
+from service.conversation_service import save_conversation,update_conversation,get_qa_by_id,get_latest_question
 from decouple import config
 import flask
 from util import load_ddl_doc
@@ -133,7 +133,7 @@ def generate_sql_2():
        data['id'] = id
        sql = data["resp"]["sql"]
        logger.info("generate sql is : "+ sql)
-        update_conversation(cvs_id, id, sql)
+        update_conversation(id, sql)
        save_save_question_async(id, user_id, question, sql)
        data["type"]="success"
        return jsonify(data)
@@ -241,7 +241,8 @@ def run_sql_2():
    logger.info("Start to run sql in main")
    try:
        id = request.args.get("id")
-        sql = get_sql_by_id(id)
+        qa = get_qa_by_id(id)
+        sql = qa["sql"]
        logger.info(f"sql is {sql}")
        if not vn.run_sql_is_set:
            return jsonify(
@@ -297,18 +298,28 @@ def query_present_question():
@app.flask_app.route("/yj_sqlbot/api/v0/gen_graph_question", methods=["GET"])
 def gen_graph_question():
    try:
-        config = {"configurable": {"thread_id": '1233'}}
+        user_id = request.args.get("user_id")
+        cvs_id = request.args.get("cvs_id")
+        config = {"configurable": {"thread_id": cvs_id}}
        question = flask.request.args.get("question")
+        question_context = get_latest_question(cvs_id, user_id,limit_count=2)
+        history = []
+        for q in question_context:
+            history.append({"role":"user",'content':q})
+        logger.info(f"history is {history}")
        initial_state: SqlAgentState = {
+            "user_id": user_id,
            "user_question": question,
-            "history": [{"role":"user",'content':'谭杰明昨日打卡记录查询'}],
+            "history": history,
            "sql_retry_count": 0,
            "chart_retry_count": 0
        }
-        result = sql_chart_agent.invoke(initial_state, config=config)
-        id = str(uuid.uuid4())
-        cache.set(id=id, field="question", value=result.get('rewritten_user_question',question))

+        result = sql_chart_agent.invoke(initial_state, config=config)
+        new_question = result.get('rewritten_user_question',question)
+        id = str(uuid.uuid4())
+        save_conversation(id, user_id, cvs_id, new_question)
+        # cache.set(id=id, field="question", value=result.get('rewritten_user_question',question))
        data = {
            'id': id,
            'sql': result.get("gen_sql_result", {}),
@@ -316,7 +327,9 @@ def gen_graph_question():
            'gen_sql_error': result.get("gen_sql_error", None),
            'gen_chart_error': result.get("gen_chart_error", None),
        }
-        cache.set(id=id, field="sql", value=data.get('sql', {}).get('sql', ''))
+        # cache.set(id=id, field="sql", value=data.get('sql', {}).get('sql', ''))
+        sql = data.get('sql', {}).get('sql', '')
+        update_conversation(id, sql)
        return jsonify(data)
    except Exception as e:
        traceback.print_exc()
@@ -325,9 +338,13 @@ def gen_graph_question():


@app.flask_app.route("/yj_sqlbot/api/v0/run_sql_3", methods=["GET"])
-@session_save
-@app.requires_cache(["sql",'question'])
-def run_sql_3(id: str, sql: str, question: str):
+# @session_save
+# @app.requires_cache(["sql",'question'])
+def run_sql_3():
+    id = request.args.get("id")
+    qa = get_qa_by_id(id)
+    sql = qa["sql"]
+    question = qa["question"]
    logger.info("Start to run sql in main")
    try:
        user_id = request.args.get("user_id")
@@ -347,6 +364,7 @@ def run_sql_3(id: str, sql: str, question: str):
        }
        config = {"configurable": {"thread_id": 'dsds'}}
        rr = result_report_agent.invoke(initial_state,config)
+        logger.info(f"rr.data is {rr.get('data', {})}")
        return jsonify(
            {
                'data': rr.get('data', {}),
--- a/service/conversation_service.py
+++ b/service/conversation_service.py
@@ -41,14 +41,14 @@ def get_conversation(cvs_id: str):
 #             }


-def update_conversation(cvs_id: str, id: str, sql=None, meta=None):
+def update_conversation(id: str, sql=None, meta=None):
    """更新sql到对应question"""
    session = SqliteSqlalchemy().session
    try:
        if sql:
-            session.query(Conversation).filter(Conversation.cvs_id == cvs_id, Conversation.id == id).update({Conversation.sql: sql})
+            session.query(Conversation).filter(Conversation.id == id).update({Conversation.sql: sql})
        if meta:
-            session.query(Conversation).filter(Conversation.cvs_id == cvs_id, Conversation.id == id).update({Conversation.meta: meta})
+            session.query(Conversation).filter(Conversation.id == id).update({Conversation.meta: meta})
        session.commit()
    except Exception as e:
        session.rollback()
@@ -72,12 +72,12 @@ def get_latest_question(cvs_id, user_id, limit_count):
        session.close()


-def get_sql_by_id(id: str):
+def get_qa_by_id(id: str):
    session = SqliteSqlalchemy().session
    try:
        result = session.query(Conversation).filter_by(id=id).first()
        if result:
-            return result.sql
+            return {"question":result.question, "sql":result.sql}
        return None
    except Exception as e:
        logger.error(f"get_sql_by_id error {e}")
--- a/service/cus_vanna_srevice.py
+++ b/service/cus_vanna_srevice.py
@@ -247,7 +247,7 @@ class OpenAICompatibleLLM(VannaBase):
                new_question = self.generate_rewritten_question(questions,**kwargs)
                logger.info(f"new_question is {new_question}")
                question = new_question if new_question else question
-                update_conversation(cvs_id, id, meta=question)
+                update_conversation(id, meta=question)

            # if user_id and cache:
            #     history = cache.get(id=user_id, field="data")
--- a/template.yaml
+++ b/template.yaml
@@ -624,4 +624,46 @@ template:
      # 最终输出
      请根据以上规则，直接输出最终生成的回复话术，不要包含任何其他解释或格式。

+  result_feedback:
+    system: |
+      # 角色 
+      你是一位经验丰富的数据分析师和SQL专家。你的核心任务是对SQL查询和其执行结果进行批判性审查。
+      # 你的任务
+      我（用户）向你提供了一系列信息，请你进行分析和反思。
+      # 输入信息
+      [原始问题]: <question>{question}</question>
+      [生成的SQL]: <sql>{sql}</sql>。
+      [执行结果]: <sql_result>{sql_result}</sql_result>。
+      [当前时间]: <current_time>{current_time}</current_time>
+      # 核心规则
+      请根据以上信息，进行全面、细致的反思，并最终判断这个结果是否能正确回答原始问题。你的反思需要包含以下几个步骤：
+      1.  **核对问题理解**：
+          回顾“原始问题”，确认其核心意图、时间范围、筛选条件和所需字段。
+          我生成的SQL是否准确捕捉了问题的所有关键要素？有没有遗漏或误解？
+      2.  **通用业务定义**
+          上周:指完整的**上周一到上周日**
+      3.  **审查SQL逻辑**：
+          逐行分析“生成的SQL”，检查其语法结构是否正确。
+          关键子句（如 `WHERE`, `GROUP BY`, `HAVING`, `JOIN`）是否准确地反映了查询意图？
+          聚合函数（如 `SUM`, `COUNT`, `AVG`）的使用是否恰当？分组维度是否正确？
+      4.  **评估结果合理性**：
+          观察“执行结果”，思考这个结果是否符合业务常识或数据的基本特征（例如，数量级、正负值、范围是否合理）？
+          结果的列名和内容是否与问题期望的输出一致？
+          如果结果为空或数据量异常，推测可能的原因。
+      5.  **最终判断和建议**：
+            **结论** : 用一句话明确指出：“结果正确”、“结果可能不正确”或“无法完全确定”。
+            **原因** : 简练地解释你做出此判断的核心理由。
+            **改进建议** (如果结果不正确):
+                         *   具体指出SQL中可能存在的逻辑错误。
+                         *   给出一个或多个修改后的SQL版本。
+                         *   解释你为什么这样修改。
+      # 输出格式
+      请严格按照以下JSON格式输出你的分析结果：
+      ```json
+      {{
+        "conclusion": "结果正确 OR 结果可能不正确 OR 无法完全确定",
+        "reasoning": "对问题理解、SQL逻辑和结果合理性的综合分析说明。",
+        "is_result_correct": true OR false,
+        "suggested_sql": "如果结论为不正确，请在此处提供修改后的SQL。如果正确，则为null。"
+      }}
 Resources: