pro-craft 0.1.22__py3-none-any.whl → 0.1.36__py3-none-any.whl

pro_craft/prompt_craft/async_.py

@@ -16,22 +16,23 @@ from datetime import datetime
 from pro_craft.utils import extract_
 import asyncio
 import re
-
+from pydantic import BaseModel, ValidationError, field_validator
 from sqlalchemy import select, desc
 from json.decoder import JSONDecodeError
+from pro_craft.database import SyncMetadata
+from datetime import datetime, timedelta
+from datetime import datetime, timedelta
+from sqlalchemy.ext.asyncio import AsyncSession, create_async_engine, async_sessionmaker
+from sqlalchemy import select, and_ # 引入 select 和 and_
+from sqlalchemy.orm import class_mapper # 用于检查对象是否是持久化的
+import tqdm
+from tqdm.asyncio import tqdm
+import pandas as pd
+import plotly.graph_objects as go
+from pro_craft import super_log
 
-class IntellectRemoveFormatError(Exception):
-    pass
-
+BATCH_SIZE = int(os.getenv("DATABASE_SYNC_BATCH_SIZE",100))
 
-def slog(s, target: str = "target",logger = None):
-    COLOR_GREEN = "\033[92m"
-    COLOR_RESET = "\033[0m" # 重置颜色
-    logger("\n"+f"{COLOR_GREEN}=={COLOR_RESET}" * 50)
-    logger(target + "\n       "+"--" * 40)
-    logger(type(s))
-    logger(s)
-    logger("\n"+f"{COLOR_GREEN}=={COLOR_RESET}" * 50)
 
 def fix_broken_json_string(broken_json_str):
     # 移除 BOM
@@ -58,6 +59,47 @@ def fix_broken_json_string(broken_json_str):
 
     return fixed_json_str
 
+async def get_last_sync_time(target_session: AsyncSession) -> datetime:
+    """从目标数据库获取上次同步时间"""
+    # 修正点：使用 select() 和 execute()
+    result = await target_session.execute(
+        select(SyncMetadata).filter_by(table_name="ai_sync_metadata")
+    )
+    metadata_entry = result.scalar_one_or_none() # 获取单个对象或 None
+
+    if metadata_entry:
+        return metadata_entry.last_sync_time
+    return datetime(1970, 1, 1) # 默认一个很早的时间
+
+async def update_last_sync_time(target_session: AsyncSession, new_sync_time: datetime):
+    """更新目标数据库的上次同步时间"""
+    # 修正点：使用 select() 和 execute()
+    result = await target_session.execute(
+        select(SyncMetadata).filter_by(table_name="ai_sync_metadata")
+    )
+    metadata_entry = result.scalar_one_or_none()
+
+    if metadata_entry:
+        metadata_entry.last_sync_time = new_sync_time
+    else:
+        # 如果不存在，则创建
+        new_metadata = SyncMetadata(table_name="ai_sync_metadata", last_sync_time=new_sync_time)
+        target_session.add(new_metadata)
+    
+    # 异步提交事务
+    await target_session.commit() # TODO
+    print(f"Updated last sync time to: {new_sync_time}")
+
+class IntellectRemoveFormatError(Exception):
+    pass
+
+class IntellectRemoveError(Exception):
+    pass
+
+class ModelNameError(Exception):
+    pass
+
+
 
 class IntellectType(Enum):
     train = "train"
@@ -76,7 +118,7 @@ class AsyncIntel():
             assert database_url
             assert 'aio' in database_url
         except AssertionError as e:
-            slog(database_url,'database_url',logger=self.logger.warning)
+            super_log(database_url,'database_url',logger=self.logger.warning)
             raise IntellectRemoveFormatError(f"异步服务url必须提供, 且必须是aiomysql配置") from e
 
         self.engine = create_async_engine(database_url, echo=False,
@@ -86,81 +128,170 @@ class AsyncIntel():
                                     pool_pre_ping=True,  # 使用前检查连接活性
                                     pool_timeout=30      # 等待连接池中连接的最长时间（秒）
                                            )
-        
 
-        if model_name in ["gemini-2.5-flash-preview-05-20-nothinking",]:
+        if "gemini" in model_name:
             self.llm = BianXieAdapter(model_name = model_name)
-        elif model_name in ["doubao-1-5-pro-256k-250115",]:
+        elif "doubao" in model_name:
             self.llm = ArkAdapter(model_name = model_name)
         else:
-            print('Use BianXieAdapter')
-            self.llm = BianXieAdapter()
+            raise ModelNameError("AsyncIntel init get error model_name from zxf")
+
+        self.eval_df = pd.DataFrame({"name":[],'status':[],"score":[],"total":[],"bad_case":[]})
             
-    async def create_database(self):
+    async def create_main_database(self):
+        tables_to_create_names = ["ai_prompts","ai_usecase"]
         async with self.engine.begin() as conn:
+            # 从 metadata 中获取对应的 Table 对象
+            specific_database_objects = []
+            for table_name in tables_to_create_names:
+                if table_name in PromptBase.metadata.tables:
+                    specific_database_objects.append(PromptBase.metadata.tables[table_name])
+                else:
+                    print(f"Warning: Table '{table_name}' not found in metadata.")
+
+            if specific_database_objects:
+                await conn.run_sync(PromptBase.metadata.create_all, tables=specific_database_objects)
+            else:
+                print("No specific tables to create.")
+
+    async def create_database(self,engine):
+        async with engine.begin() as conn:
             await conn.run_sync(PromptBase.metadata.create_all)
-            
-    async def _get_latest_prompt_version(self,target_prompt_id,session):
+
+    async def get_prompt(self,prompt_id,version,session):
         """
         获取指定 prompt_id 的最新版本数据，通过创建时间判断。
         """
-        stmt = select(Prompt).filter(
-            Prompt.prompt_id == target_prompt_id
-        ).order_by(
-            desc(Prompt.timestamp), # 使用 sqlalchemy.desc() 来指定降序
-            desc(Prompt.version)    # 使用 sqlalchemy.desc() 来指定降序
-        )
-        
+        if version:
+            stmt_ = select(Prompt).filter(
+                Prompt.prompt_id == prompt_id,
+                Prompt.version == version
+            )
+        else:  
+            stmt_ = select(Prompt).filter(
+                Prompt.prompt_id == prompt_id,
+            )
+        stmt = stmt_.order_by(
+                desc(Prompt.timestamp), # 使用 sqlalchemy.desc() 来指定降序
+                desc(Prompt.version)    # 使用 sqlalchemy.desc() 来指定降序
+            )
+
         result = await session.execute(stmt)
-        # 3. 从 Result 对象中获取第一个模型实例
-        # .scalars() 用于从结果行中获取第一个列的值（这里是Prompt对象本身）
-        # .first() 获取第一个结果
         result = result.scalars().first()
 
         return result
 
-    async def _get_specific_prompt_version(self,target_prompt_id, target_version,session):
-        """
-        获取指定 prompt_id 和特定版本的数据。
+    async def sync_production_database(self,database_url:str):
+        target_engine = create_async_engine(database_url, echo=False)
+        await self.create_database(target_engine) 
+        async with create_async_session(self.engine) as source_session:
+            async with create_async_session(target_engine) as target_session:
+            
+                last_sync_time = await get_last_sync_time(target_session)
+                print(f"Starting sync for sync_metadata from: {last_sync_time}")
+
+
+                processed_count = 0
+                #2 next_sync_watermark = last_sync_time
+                current_batch_max_updated_at = last_sync_time
+
+                while True:
+                    source_results = await source_session.execute(
+                        select(Prompt)
+                        .filter(Prompt.timestamp > last_sync_time)
+                        .order_by(Prompt.timestamp.asc(), Prompt.id.asc())
+                        .limit(BATCH_SIZE)
+                    )
+                    records_to_sync = source_results.scalars().all()
+                    if not records_to_sync:
+                        print("没有更多记录了")
+                        break # 没有更多记录了
+
+                    #2 max_timestamp_in_batch = datetime(1970, 1, 1) # 初始化为最早时间
+
+                    # 准备要插入或更新到目标数据库的数据
+                    for record in records_to_sync:
+                        # 查找目标数据库中是否存在该ID的记录
+                        # 这里的 `User` 模型会对应到 target_db.users
+                        target_prompt_result = await target_session.execute(
+                            select(Prompt).filter_by(id=record.id) # 假设 prompt_id 是唯一标识符
+                        )
+                        target_prompt = target_prompt_result.scalar_one_or_none()
+                        
+                        if target_prompt:
+                            # 如果存在，则更新
+                            target_prompt.prompt_id = record.prompt_id
+                            target_prompt.version = record.version
+                            target_prompt.timestamp = record.timestamp
+                            target_prompt.prompt = record.prompt
+                            target_prompt.use_case = record.use_case
+                            target_prompt.action_type = record.action_type
+                            target_prompt.demand = record.demand
+                            target_prompt.score = record.score
+                            target_prompt.is_deleted = record.is_deleted
+                        else:
+                            # 如果不存在，则添加新记录
+                            # 注意：这里需要创建一个新的User实例，而不是直接添加源数据库的record对象
+                            new_prompt = Prompt(
+                                prompt_id=record.prompt_id, 
+                                version=record.version,
+                                timestamp=record.timestamp,
+                                prompt = record.prompt,
+                                use_case = record.use_case,
+                                action_type = record.action_type,
+                                demand = record.demand,
+                                score = record.score,
+                                is_deleted = record.is_deleted
+                                )
+                            target_session.add(new_prompt)
+                        
+                        # 记录当前批次最大的 updated_at
+                        #2 
+                        # if record.timestamp > max_timestamp_in_batch:
+                        #     max_timestamp_in_batch = record.timestamp
+                        if record.timestamp > current_batch_max_updated_at:
+                            current_batch_max_updated_at = record.timestamp
+
+
+                    await target_session.commit() 
+                    processed_count += len(records_to_sync)
+                    print(f"Processed {len(records_to_sync)} records. Total processed: {processed_count}")
+
+                    #2 next_sync_watermark = max_timestamp_in_batch + timedelta(microseconds=1)
+                    last_sync_time = current_batch_max_updated_at + timedelta(microseconds=1) 
 
-        Args:
-            target_prompt_id (str): 目标提示词的唯一标识符。
-            target_version (int): 目标提示词的版本号。
-            table_name (str): 存储提示词数据的数据库表名。
-            db_manager (DBManager): 数据库管理器的实例，用于执行查询。
+                    
+                    if len(records_to_sync) < BATCH_SIZE: # 如果查询到的记录数小于批次大小，说明已经处理完所有符合条件的记录
+                        break
 
-        Returns:
-            dict or None: 如果找到，返回包含 id, prompt_id, version, timestamp, prompt 字段的字典；
-                        否则返回 None。
-        """
-        stmt = select(Prompt).filter(
-        Prompt.prompt_id == target_prompt_id,
-        Prompt.version == target_version
-        )
-        result = await session.execute(stmt)
+                if processed_count > 0:
+                    # 最终更新last_sync_time到数据库，确保记录的是所有已处理记录中最新的一个
+                    await update_last_sync_time(target_session, current_batch_max_updated_at + timedelta(microseconds=1))
+
+                    #2 await update_last_sync_time(target_session, next_sync_watermark)
 
-        specific_prompt = result.scalars().one_or_none()
+                    await target_session.commit() # 确保最终的 metadata 更新也被提交
+                else:
+                    print("No new records to sync.")
 
-        return specific_prompt
 
-    async def get_prompts_from_sql(self,
+    async def get_prompt_safe(self,
                              prompt_id: str,
                              version = None,
                              session = None) -> Prompt:
         """
         从sql获取提示词
         """
-        # 查看是否已经存在
-        if version:
-            prompts_obj = await self._get_specific_prompt_version(prompt_id,version,session=session)
-            if not prompts_obj:
-                prompts_obj = await self._get_latest_prompt_version(prompt_id,session = session)
-        else:
-            prompts_obj = await self._get_latest_prompt_version(prompt_id,session = session)     
-        return prompts_obj
+        prompt_obj = await self.get_prompt(prompt_id=prompt_id,version=version,session=session)
+        if prompt_obj:
+            return prompt_obj
+
+        print("warnning 未找到制定版本, 默认使用最新版本")
+        prompt_obj = await self.get_prompt(prompt_id=prompt_id,version=None,session=session)
+        return prompt_obj
+
         
-            
-    async def save_prompt_increment_version(self,
+    async def save_prompt(self,
                            prompt_id: str,
                            new_prompt: str,
                            use_case:str = "",
@@ -173,7 +304,7 @@ class AsyncIntel():
         input_data 指的是输入用例, 可以为空
         """
         # 查看是否已经存在
-        prompts_obj = await self.get_prompts_from_sql(prompt_id=prompt_id,session=session)
+        prompts_obj = await self.get_prompt_safe(prompt_id=prompt_id,session=session)
 
         if prompts_obj:
             # 如果存在版本加1
@@ -200,105 +331,55 @@ class AsyncIntel():
         session.add(prompt1)
         await session.commit() # 提交事务，将数据写入数据库
 
-    async def save_use_case_by_sql(self,
+    async def get_use_case(self,
+                             target_prompt_id: str,
+                             session = None
+                            ):
+        """
+        从sql保存提示词
+        """
+        stmt = select(UseCase).filter(UseCase.is_deleted == 0,
+                                      UseCase.prompt_id == target_prompt_id)
+        
+        result = await session.execute(stmt)
+        # use_case = result.scalars().one_or_none()
+        use_case = result.scalars().all()
+        return use_case
+
+    async def save_use_case(self,
                              prompt_id: str,
                              use_case:str = "",
+                             timestamp = "",
                              output = "",
                              solution: str = "",
+                             faired_time  = 0,
                              session = None
                             ):
+
         """
         从sql保存提示词
         """
-        use_case = UseCase(prompt_id=prompt_id, 
+        #TODO 存之前保证数据库中相同的prompt_id中没有重复的use_case
+        use_cases = await self.get_use_case(target_prompt_id = prompt_id,
+                                  session = session)
+        for use_case_old in use_cases:
+            if use_case == use_case_old.use_case:
+                print("用例已经存在")
+                return 
+        
+        use_case = UseCase(prompt_id=prompt_id,
                         use_case = use_case,
+                        timestamp = timestamp,
                         output = output,
                         solution = solution,
+                        faired_time = faired_time,
                         )
 
         session.add(use_case)
         await session.commit() # 提交事务，将数据写入数据库
 
-    async def summary_to_sql(
-            self,
-            prompt_id:str,
-            version = None,
-            prompt = "",
-            session = None
-        ):
-        """
-        让大模型微调已经存在的 system_prompt
-        """
-        system_prompt_created_prompt = """        
-很棒, 我们已经达成了某种默契, 我们之间合作无间, 但是, 可悲的是, 当我关闭这个窗口的时候, 你就会忘记我们之间经历的种种磨合, 这是可惜且心痛的, 所以你能否将目前这一套处理流程结晶成一个优质的prompt 这样, 我们下一次只要将prompt输入, 你就能想起我们今天的磨合过程,
-对了,我提示一点, 这个prompt的主角是你, 也就是说, 你在和未来的你对话, 你要教会未来的你今天这件事, 是否让我看懂到时其次
-
-只要输出提示词内容即可, 不需要任何的说明和解释
-"""
-        system_result = await self.llm.aproduct(prompt + system_prompt_created_prompt)
-
-        s_prompt = extract_(system_result,pattern_key=r"prompt")
-        chat_history = s_prompt or system_result
-        await self.save_prompt_increment_version(prompt_id,
-                                new_prompt = chat_history,
-                                use_case = " summary ",
-                                score = 60,
-                                session = session)
-        
-    async def prompt_finetune_to_sql(
-            self,
-            prompt_id:str,
-            version = None,
-            demand: str = "",
-            session = None,
-        ):
-        """
-        让大模型微调已经存在的 system_prompt
-        """
-        change_by_opinion_prompt = """
-你是一个资深AI提示词工程师，具备卓越的Prompt设计与优化能力。
-我将为你提供一段现有System Prompt。你的核心任务是基于这段Prompt进行修改，以实现我提出的特定目标和功能需求。
-请你绝对严格地遵循以下原则：
- 极端最小化修改原则（核心）：
- 在满足所有功能需求的前提下，只进行我明确要求的修改。
- 即使你认为有更“优化”、“清晰”或“简洁”的表达方式，只要我没有明确要求，也绝不允许进行任何未经指令的修改。
- 目的就是尽可能地保留原有Prompt的字符和结构不变，除非我的功能要求必须改变。
- 例如，如果我只要求你修改一个词，你就不应该修改整句话的结构。
- 严格遵循我的指令：
- 你必须精确地执行我提出的所有具体任务和要求。
- 绝不允许自行添加任何超出指令范围的说明、角色扮演、约束条件或任何非我指令要求的内容。
- 保持原有Prompt的风格和语调：
- 尽可能地与现有Prompt的语言风格、正式程度和语调保持一致。
- 不要改变不相关的句子或其表达方式。
- 只提供修改后的Prompt：
- 直接输出修改后的完整System Prompt文本。
- 不要包含任何解释、说明或额外对话。
- 在你开始之前，请务必确认你已理解并能绝对严格地遵守这些原则。任何未经明确指令的改动都将视为未能完成任务。
-
-现有System Prompt:
-{old_system_prompt}
-
-功能需求:
-{opinion}
-"""
-
-        prompt_ = await self.get_prompts_from_sql(prompt_id = prompt_id,version = version,
-                                                    session=session)
-        if demand:
-            new_prompt = await self.llm.aproduct(
-                change_by_opinion_prompt.format(old_system_prompt=prompt_.prompt, opinion=demand)
-            )
-        else:
-            new_prompt = prompt_
-        await self.save_prompt_increment_version(prompt_id = prompt_id,
-                            new_prompt = new_prompt,
-                            use_case = " finetune ",
-                            score = 60,
-                            session = session)
-
-
     async def push_action_order(self,demand : str,prompt_id: str,
-                         action_type = 'train'):
+                         action_type = 'train'):# init
 
         """
         从sql保存提示词
@@ -309,9 +390,9 @@ class AsyncIntel():
         # 查看是否已经存在
         async with create_async_session(self.engine) as session:
 
-            latest_prompt = await self.get_prompts_from_sql(prompt_id=prompt_id,session=session)
-
-            await self.save_prompt_increment_version(prompt_id=latest_prompt.prompt_id,
+            latest_prompt = await self.get_prompt_safe(prompt_id=prompt_id,session=session)
+            if latest_prompt:
+                await self.save_prompt(prompt_id=latest_prompt.prompt_id,
                                     new_prompt = latest_prompt.prompt,
                                     use_case = latest_prompt.use_case,
                                     action_type=action_type,
@@ -319,17 +400,26 @@ class AsyncIntel():
                                     score=latest_prompt.score,
                                     session=session
                                     )
-            
-            return "success"
-
+                return "success"
+            else:
+                await self.save_prompt(prompt_id=prompt_id,
+                                    new_prompt = demand,
+                                    use_case = "",
+                                    action_type="inference",
+                                    demand=demand,
+                                    score=60,
+                                    session=session
+                                    )
+                return "init"
 
 
-    async def intellect_remove(self,
+    async def intellect(self,
                     input_data: dict | str,
                     output_format: str,
                     prompt_id: str,
                     version: str = None,
                     inference_save_case = True,
+                    change_case = False,
                     ):
         if isinstance(input_data,dict):
             input_ = json.dumps(input_data,ensure_ascii=False)
@@ -338,29 +428,22 @@ class AsyncIntel():
         
         # 查数据库, 获取最新提示词对象
         async with create_async_session(self.engine) as session:
-            result_obj = await self.get_prompts_from_sql(prompt_id=prompt_id,session=session)
-
-
+            result_obj = await self.get_prompt_safe(prompt_id=prompt_id,session=session)
             if result_obj is None:
-                await self.save_prompt_increment_version(
-                    prompt_id = prompt_id,
-                    new_prompt = "做一些处理",
-                    use_case = input_,
-                    score = 60,
-                    session = session
-                )
-                ai_result = "初始化完成"
-                return ai_result
+                raise IntellectRemoveError("不存在的prompt_id")
 
             prompt = result_obj.prompt
             if result_obj.action_type == "inference":
                 # 直接推理即可
-                ai_result = await self.llm.aproduct(prompt + output_format + "\n-----input----\n" +  input_)
+                ai_result = await self.llm.aproduct(prompt + output_format + "\nuser:" +  input_)
                 if inference_save_case:
-                    await self.save_use_case_by_sql(prompt_id,
+                    # 设计一个机制, 传输说获取300数据, 那么数据就一直流转获取, 知道300截止
+                    await self.save_use_case(prompt_id,
                                         use_case = input_,
+                                        timestamp = datetime.now(),
                                         output = ai_result,
-                                        solution = "备注/理想回复",
+                                        solution = output_format,
+                                        faired_time = 0,
                                         session = session,
                                         )
                     
@@ -369,9 +452,6 @@ class AsyncIntel():
                 # 则训练推广
 
                 # 新版本 默人修改会 inference 状态
-                chat_history = prompt
-                before_input = result_obj.use_case
-                demand = result_obj.demand
             
 
                 # assert demand
@@ -386,195 +466,170 @@ class AsyncIntel():
             
                 # ai_result = await self.llm.aproduct(input_prompt)
                 # chat_history = input_prompt + "\nassistant:\n" + ai_result # 用聊天记录作为完整提示词
-                # await self.save_prompt_increment_version(prompt_id, chat_history,
+                # await self.save_prompt(prompt_id, chat_history,
                 #                         use_case = input_,
                 #                         score = 60,
                 #                         session = session)
 
-                if input_ == before_input:
-                    new_prompt = prompt + "\nuser:" + demand
-                else:
-                    new_prompt = prompt + "\nuser:" + input_
+
+                # version 2
+
+                # if input_ == before_input:
+                #     new_prompt = prompt + "\nuser:" + demand
+                # else:
+                #     new_prompt = prompt + "\nuser:" + input_
+
+                # ai_result = await self.llm.aproduct(new_prompt + output_format)
+
+                # save_new_prompt = new_prompt + "\nassistant:\n" + ai_result
+
+
+                # await self.save_prompt(
+                #     prompt_id, 
+                #     new_prompt=save_new_prompt,
+                #     use_case = input_,
+                #     action_type = "inference",
+                #     score = 60,
+                #     session = session)
+                chat_history = prompt
+                before_input = result_obj.use_case
+                demand = result_obj.demand
+                input_data = input_
+                if before_input == "" or change_case is True:
+                    result_obj.use_case = input_
+                    await session.commit()
+                    # 查询上一条, 将before_input 更新位input_
+                    prompt += input_
+
+                # 使用更新后的数据进行后续步骤
+                new_prompt = prompt + "\nuser:" + demand
 
                 ai_result = await self.llm.aproduct(new_prompt + output_format)
 
                 save_new_prompt = new_prompt + "\nassistant:\n" + ai_result
 
 
-                await self.save_prompt_increment_version(
+                await self.save_prompt(
                     prompt_id, 
                     new_prompt=save_new_prompt,
                     use_case = input_,
+                    action_type = "inference",
                     score = 60,
                     session = session)
     
             elif result_obj.action_type == "summary":
+                system_prompt_summary = """        
+很棒, 我们已经达成了某种默契, 我们之间合作无间, 但是, 可悲的是, 当我关闭这个窗口的时候, 你就会忘记我们之间经历的种种磨合, 这是可惜且心痛的, 所以你能否将目前这一套处理流程结晶成一个优质的prompt 这样, 我们下一次只要将prompt输入, 你就能想起我们今天的磨合过程,
+对了,我提示一点, 这个prompt的主角是你, 也就是说, 你在和未来的你对话, 你要教会未来的你今天这件事, 是否让我看懂到时其次
 
-                await self.summary_to_sql(prompt_id = prompt_id,
-                            prompt = prompt,
-                            session = session
-                            )
-                ai_result = await self.llm.aproduct(prompt + output_format + "\n-----input----\n" +  input_)
+只要输出提示词内容即可, 不需要任何的说明和解释
+"""
+                
+                latest_prompt = await self.get_prompt_safe(prompt_id=prompt_id,session=session)
+
+                system_result = await self.llm.aproduct(prompt + system_prompt_summary)
+                s_prompt = extract_(system_result,pattern_key=r"prompt")
+                new_prompt = s_prompt or system_result
+                await self.save_prompt(
+                    prompt_id,
+                    new_prompt = new_prompt,
+                    use_case = latest_prompt.use_case,
+                    score = 65,
+                    action_type = "inference",
+                    session = session
+                    )
+
+                ai_result = await self.llm.aproduct(prompt + output_format + "\nuser:" +  input_)
 
             elif result_obj.action_type == "finetune":
                 demand = result_obj.demand
-            
-                assert demand
-                await self.prompt_finetune_to_sql(prompt_id = prompt_id,
-                                            demand = demand,
-                                            session = session
-                                            )
-                ai_result = await self.llm.aproduct(prompt + output_format + "\n-----input----\n" +  input_)
-            elif result_obj.action_type == "patch":
-                demand = result_obj.demand
-                assert demand
-                chat_history = prompt + demand
-                ai_result = await self.llm.aproduct(chat_history + output_format + "\n-----input----\n" +  input_)
-                self.save_prompt_increment_version(prompt_id, 
-                                                   chat_history,
-                                                    use_case = input_,
-                                                    score = 60,
-                                                    session = session)
-
-            else:
-                raise
-
-        return ai_result
-    
-    async def intellect_stream_remove(self,
-                    input_data: dict | str,
-                    output_format: str,
-                    prompt_id: str,
-                    version: str = None,
-                    inference_save_case = True,
-                    push_patch = False,
-                    ):
-        if isinstance(input_data,dict):
-            input_ = json.dumps(input_data,ensure_ascii=False)
-        elif isinstance(input_data,str):
-            input_ = input_data
-
-        
-        # 查数据库, 获取最新提示词对象
-        with create_session(self.engine) as session:
-            result_obj = await self.get_prompts_from_sql(prompt_id=prompt_id,session=session)
-
-            '''
-                        if result_obj is None:
-                            await self.save_prompt_increment_version(
-                                prompt_id = prompt_id,
-                                new_prompt = "做一些处理",
-                                use_case = input_,
-                                session = session
-                            )
-                            ai_result = await self.intellect_stream_remove(input_data = input_data,
-                                                output_format = output_format,
-                                                prompt_id = prompt_id,
-                                                version = version,
-                                                inference_save_case = inference_save_case
-                                                )
-                            return ai_result'''
+                change_by_opinion_prompt = """
+你是一个资深AI提示词工程师，具备卓越的Prompt设计与优化能力。
+我将为你提供一段现有System Prompt。你的核心任务是基于这段Prompt进行修改，以实现我提出的特定目标和功能需求。
+请你绝对严格地遵循以下原则：
+极端最小化修改原则（核心）：
+在满足所有功能需求的前提下，只进行我明确要求的修改。
+即使你认为有更“优化”、“清晰”或“简洁”的表达方式，只要我没有明确要求，也绝不允许进行任何未经指令的修改。
+目的就是尽可能地保留原有Prompt的字符和结构不变，除非我的功能要求必须改变。
+例如，如果我只要求你修改一个词，你就不应该修改整句话的结构。
+严格遵循我的指令：
+你必须精确地执行我提出的所有具体任务和要求。
+绝不允许自行添加任何超出指令范围的说明、角色扮演、约束条件或任何非我指令要求的内容。
+保持原有Prompt的风格和语调：
+尽可能地与现有Prompt的语言风格、正式程度和语调保持一致。
+不要改变不相关的句子或其表达方式。
+只提供修改后的Prompt：
+直接输出修改后的完整System Prompt文本。
+不要包含任何解释、说明或额外对话。
+在你开始之前，请务必确认你已理解并能绝对严格地遵守这些原则。任何未经明确指令的改动都将视为未能完成任务。
 
-            prompt = result_obj.prompt
-            if result_obj.action_type == "inference":
-                # 直接推理即可
-                
-                ai_generate_result = self.llm.aproduct_stream(prompt + output_format +  "\n-----input----\n" +  input_)
-                ai_result = ""
-                async for word in ai_generate_result:
-                    ai_result += word
-                    yield word
-                if inference_save_case:
-                    await self.save_use_case_by_sql(prompt_id,
-                                        use_case = input_,
-                                        output = ai_result,
-                                        solution = "备注/理想回复",
-                                        session = session,
-                                        )
-                    
-            elif result_obj.action_type == "train":
-                assert result_obj.demand # 如果type = train 且 demand 是空 则报错
-                # 则训练推广
+现有System Prompt:
+{old_system_prompt}
 
-                # 新版本 默人修改会 inference 状态
-                chat_history = prompt
-                before_input = result_obj.use_case
-                demand = result_obj.demand
-            
+功能需求:
+{opinion}
+"""
 
+                latest_prompt = await self.get_prompt_safe(prompt_id=prompt_id,session=session)
+                prompt_ = await self.get_prompt_safe(prompt_id = prompt_id,version = version,
+                                                                session=session)
                 assert demand
-                # 注意, 这里的调整要求使用最初的那个输入, 最好一口气调整好
-                chat_history = prompt
-                if input_ == before_input: # 输入没变, 说明还是针对同一个输入进行讨论
-                    # input_prompt = chat_history + "\nuser:" + demand
-                    input_prompt = chat_history + "\nuser:" + demand + output_format 
+                
+                if demand:
+                        new_prompt = await self.llm.aproduct(
+                            change_by_opinion_prompt.format(old_system_prompt=prompt_.prompt, opinion=demand)
+                        )
                 else:
-                    # input_prompt = chat_history + "\nuser:" + demand + "\n-----input----\n" + input_
-                    input_prompt = chat_history + "\nuser:" + demand + output_format  + "\n-----input----\n" + input_
-            
-                ai_generate_result = self.llm.aproduct_stream(input_prompt)
-                ai_result = ""
-                async for word in ai_generate_result:
-                    ai_result += word
-                    yield word
-
-                chat_history = input_prompt + "\nassistant:\n" + ai_result # 用聊天记录作为完整提示词
-                await self.save_prompt_increment_version(prompt_id, chat_history,
-                                        use_case = input_,
-                                        score = 60,
-                                        session = session)
-
-    
-            elif result_obj.action_type == "summary":
-
-                await self.summary_to_sql(prompt_id = prompt_id,
-                            prompt = prompt,
-                            session = session
-                            )
-                input_prompt = prompt + output_format + "\n-----input----\n" +  input_
-                ai_generate_result = self.llm.aproduct_stream(input_prompt)
-                ai_result = ""
-                async for word in ai_generate_result:
-                    ai_result += word
-                    yield word
+                    new_prompt = prompt_ 
+                await self.save_prompt(
+                    prompt_id,
+                    new_prompt = new_prompt,
+                    use_case = latest_prompt.use_case,
+                    score = 70,
+                    action_type = "inference",
+                    session = session
+                    )
                 
-            elif result_obj.action_type == "finetune":
-                demand = result_obj.demand
-            
-                assert demand
-                await self.prompt_finetune_to_sql(prompt_id = prompt_id,
-                                            demand = demand,
-                                            session = session
-                                            )
-                input_prompt = prompt + output_format + "\n-----input----\n" +  input_
-                ai_generate_result = self.llm.aproduct_stream(input_prompt)
-                ai_result = ""
-                async for word in ai_generate_result:
-                    ai_result += word
-                    yield word
+                ai_result = await self.llm.aproduct(prompt + output_format + "\nuser:" +  input_)
 
             elif result_obj.action_type == "patch":
-                
                 demand = result_obj.demand
                 assert demand
-
+                latest_prompt = await self.get_prompt_safe(prompt_id=prompt_id,session=session)
+                
                 chat_history = prompt + demand
-                ai_generate_result = self.llm.aproduct_stream(chat_history + output_format + "\n-----input----\n" +  input_)
-                ai_result = ""
-                async for word in ai_generate_result:
-                    ai_result += word
-                    yield word
-                if push_patch:
-                    self.save_prompt_increment_version(prompt_id, chat_history,
-                                            use_case = input_,
-                                            score = 60,
-                                            session = session)
+                await self.save_prompt(prompt_id, 
+                                chat_history,
+                                use_case = latest_prompt.use_case,
+                                score = 70,
+                                action_type = "inference",
+                                session = session)
+                
+                ai_result = await self.llm.aproduct(chat_history + output_format + "\nuser:" +  input_)
+                
+            elif result_obj.action_type.startswith("to:"):
+                target_version = result_obj.action_type.split(":")[-1]
+                latest_prompt = await self.get_prompt_safe(prompt_id=prompt_id,session=session)
+                prompt_obj = await self.get_prompt_safe(prompt_id=prompt_id,
+                                     version=target_version,
+                                     session=session)
+                
+                await self.save_prompt(prompt_id, 
+                                prompt_obj.prompt,
+                                use_case = latest_prompt.use_case,
+                                score = prompt_obj.score,
+                                action_type = "inference",
+                                session = session)
+                ai_result = await self.llm.aproduct(prompt_obj.prompt + output_format + "\nuser:" +  input_)
+                
             else:
                 raise
 
-    async def intellect_remove_format(self,
+        return ai_result
+        
+    async def intellect_format(self,
                     input_data: dict | str,
-                    OutputFormat: object,
+                    OutputFormat: object | None,
                     prompt_id: str,
                     ExtraFormats: list[object] = [],
                     version: str = None,
@@ -588,9 +643,11 @@ class AsyncIntel():
 "```json([\s\S]*?)```"
 使用以下方式验证
 """
-        output_format = base_format_prompt + "\n".join([inspect.getsource(outputformat) for outputformat in ExtraFormats]) + inspect.getsource(OutputFormat)
-
-        ai_result = await self.intellect_remove(
+        if OutputFormat:
+            output_format = base_format_prompt + "\n".join([inspect.getsource(outputformat) for outputformat in ExtraFormats]) + inspect.getsource(OutputFormat)
+        else:
+            output_format = ""
+        ai_result = await self.intellect(
                     input_data=input_data,
                     output_format=output_format,
                     prompt_id=prompt_id,
@@ -599,26 +656,32 @@ class AsyncIntel():
                 )
 
         try:
-
             json_str = extract_(ai_result,r'json')
-            # json_str = fix_broken_json_string(json_str)
             ai_result = json.loads(json_str)
-            OutputFormat(**ai_result)
+            if OutputFormat:
+                OutputFormat(**ai_result)
 
         except JSONDecodeError as e:
-            slog(ai_result,logger=self.logger.error)
             try:
                 self.logger.error(f"尝试补救")
                 json_str = fix_broken_json_string(json_str)
                 ai_result = json.loads(json_str)
-                OutputFormat(**ai_result)
+                if OutputFormat:
+                    OutputFormat(**ai_result)
 
             except JSONDecodeError as e:
-                raise IntellectRemoveFormatError(f"prompt_id: {prompt_id} 在生成后做json解析时报错") from e
+                raise IntellectRemoveFormatError(f"prompt_id: {prompt_id} 生成的内容为无法被Json解析 {e}") from e
+        
+        except ValidationError as e:
+            err_info = e.errors()[0]
+            raise IntellectRemoveFormatError(f"{err_info["type"]}: 属性:{err_info['loc']}, 发生了如下错误: {err_info['msg']}, 格式校验失败, 当前输入为: {err_info['input']} 请检查") from e
+
+        except Exception as e:
+            raise Exception(f"Error {prompt_id} : {e}") from e
 
         return ai_result
     
-    async def intellect_remove_formats(self,
+    async def intellect_formats(self,
                     input_datas: list[dict | str],
                     OutputFormat: object,
                     prompt_id: str,
@@ -628,14 +691,16 @@ class AsyncIntel():
                     ):
                 
         async with create_async_session(self.engine) as session:
-            prompt_result = await self.get_prompts_from_sql(prompt_id=prompt_id,
+            prompt_result = await self.get_prompt_safe(prompt_id=prompt_id,
                                                                    session=session)
+            if prompt_result is None:
+                raise IntellectRemoveError("不存在的prompt_id")
         if prompt_result.action_type != "inference":
             input_datas = input_datas[:1]
         tasks = []
         for input_data in input_datas:
             tasks.append(
-                self.intellect_remove_format(
+                self.intellect_format(
                     input_data = input_data,
                     prompt_id = prompt_id,
                     OutputFormat = OutputFormat,
@@ -646,9 +711,8 @@ class AsyncIntel():
             )
         results = await asyncio.gather(*tasks, return_exceptions=False)
         return results
-
     
-    def intellect_remove_warp(self,prompt_id: str):
+    def intellect_warp(self,prompt_id: str):
         def outer_packing(func):
             @functools.wraps(func)
             async def wrapper(*args, **kwargs):
@@ -663,7 +727,7 @@ class AsyncIntel():
                 elif isinstance(input_data,str):
                     input_ = output_ = input_data
 
-                output_ = await self.intellect_remove_format(
+                output_ = await self.intellect_format(
                         input_data = input_data,
                         prompt_id = prompt_id,
                         OutputFormat = OutputFormat,
@@ -675,3 +739,182 @@ class AsyncIntel():
                 return result
             return wrapper
         return outer_packing
+
+    async def intellect_format_eval(self,
+                    OutputFormat: object,
+                    prompt_id: str,
+                    ExtraFormats: list[object] = [],
+                    version: str = None,
+                    MIN_SUCCESS_RATE = 80.0,
+                    ConTent_Function = None,
+                    ):
+        """
+        ConTent_Function: 
+        # TODO 人类评价 eval
+        # TODO llm 评价 eval
+        """
+        async with create_async_session(self.engine) as session:
+            use_cases = await self.get_use_case(target_prompt_id=prompt_id,session=session)
+            prompt_result = await self.get_prompt_safe(prompt_id=prompt_id,
+                                                                   session=session)
+            if prompt_result is None:
+                raise IntellectRemoveError("不存在的prompt_id")
+            if prompt_result.action_type != "inference":
+                raise IntellectRemoveError("请在inference模式下使用次类")
+            
+
+            total_assertions = len(use_cases)
+            result_cases = []
+
+            async def evals_func(use_case,prompt_id,OutputFormat,ExtraFormats,version):
+                try:
+                    # 这里将参数传入
+                    ai_result = await self.intellect_format(
+                        input_data = use_case.use_case,
+                        prompt_id = prompt_id,
+                        OutputFormat = OutputFormat,
+                        ExtraFormats = ExtraFormats,
+                        version = version,
+                        inference_save_case = False,
+                    )
+                    if ConTent_Function:
+                        ConTent_Function()
+                    result_cases.append({"type":"Successful","case":use_case.use_case,"reply":f"pass"})
+                    use_case.output = ai_result
+                except IntellectRemoveFormatError as e:
+                    result_cases.append({"type":"FAILED","case":use_case.use_case,"reply":f"{e}"})
+                    use_case.output = f"{"FAILED"}-{e}"
+                except Exception as e: # 捕获其他可能的错误
+                    result_cases.append({"type":"FAILED","case":use_case.use_case,"reply":f"Exp {e}"})
+                    use_case.output = f"{"FAILED"}-{e}"
+                await session.commit()
+
+            tasks = []
+            for use_case in use_cases:
+                tasks.append(
+                    evals_func(
+                        use_case = use_case,
+                        prompt_id = prompt_id,
+                        OutputFormat = OutputFormat,
+                        ExtraFormats = ExtraFormats,
+                        version = version
+                    )
+                )
+            await tqdm.gather(*tasks,total=len(tasks))
+            # await asyncio.gather(*tasks, return_exceptions=False)
+
+
+            successful_assertions = 0
+            bad_case = []
+            for i in result_cases:
+                if i['type'] == "Successful":
+                    successful_assertions += 1
+                else:
+                    bad_case.append(i)
+
+            success_rate = (successful_assertions / total_assertions) * 100
+
+            if success_rate >= MIN_SUCCESS_RATE:
+                return "通过", success_rate, str(total_assertions), json.dumps(bad_case,ensure_ascii=False),
+            else:
+                return "未通过",success_rate, str(total_assertions), json.dumps(bad_case,ensure_ascii=False),
+
+
+
+    def draw_data(self):
+        df = self.eval_df
+        # --- 可视化部分 ---
+        fig = go.Figure()
+
+        # 为每个条形图动态设置颜色
+        colors = []
+        for status_val in df['status']:
+            if status_val == '通过':
+                colors.append('mediumseagreen') # 通过为绿色
+            else: # 假设其他所有状态都视为“未通过”
+                colors.append('lightcoral') # 未通过为红色
+
+        fig.add_trace(go.Bar(
+            y=df['name'], # Y轴显示项目名称
+            x=df['score'],       # X轴显示通过百分比 (score列现在代表通过百分比)
+            orientation='h',     # 设置为横向
+            name='通过率',       # 这个 name 可能会在图例中显示
+            marker_color=colors, # !!! 这里根据 status 动态设置颜色 !!!
+            text=df['score'].apply(lambda x: f'{x:.2f}%'), # 在条形图上显示百分比文本
+            textposition='inside',
+            insidetextanchor='middle',
+            hovertemplate="<b>prompt:</b> %{y}<br><b>状态:</b> " + df['status'] + "<br><b>总量:</b> "+ df['total'] + "<br><b>通过百分比:</b> %{x:.2f}%<extra></extra>"
+        ))
+
+        # 添加一个辅助的条形图作为背景，表示总的100%
+        fig.add_trace(go.Bar(
+            y=df['name'],
+            x=[100] * len(df), # 所有项目都填充到100%
+            orientation='h',
+            name='总计',
+            marker_color='lightgray', # 背景用灰色
+            hoverinfo='none', # 不显示hover信息
+            opacity=0.5, # 设置透明度
+            showlegend=False # 不显示图例
+        ))
+
+        fig.update_layout(
+            title='各项目/批次通过百分比及状态',
+            xaxis=dict(
+                title='通过百分比 (%)',
+                range=[0, 100], # X轴范围0-100
+                tickvals=[0, 25, 50, 75, 100],
+                showgrid=True,
+                gridcolor='lightgray'
+            ),
+            yaxis=dict(
+                title='项目/批次',
+                autorange="reversed"
+            ),
+            barmode='overlay', # 仍使用 overlay 模式，因为背景条是独立的
+            hovermode="y unified",
+            margin=dict(l=100, r=20, t=60, b=50),
+            height=400 + len(df) * 30
+        )
+        error_message =str(df['bad_case'].to_dict())
+        fig.add_annotation(
+            text=f"<b>bad_case:</b> {error_message}", # 要显示的文本
+            xref="paper", yref="paper", # 使用“paper”坐标系，表示相对于图表区域
+            x=0.01, y=-0.15, # x=0.01 靠近左侧，y=-0.15 在图表底部下方 (您可以调整这些值)
+            showarrow=False, # 不显示箭头
+            align="left",
+            font=dict(
+                family="Arial, sans-serif",
+                size=12,
+                color="red" # 错误信息通常用红色
+            ),
+            bgcolor="white", # 背景颜色
+            bordercolor="red", # 边框颜色
+            borderwidth=1,
+            borderpad=4,
+            xanchor='left', # 文本框左对齐到x坐标
+            yanchor='top' # 文本框顶部对齐到y坐标
+        )
+        # 可能还需要调整底部的边距以容纳错误信息
+        fig.update_layout(
+            margin=dict(l=100, r=20, t=60, b=100), # 增加底部边距
+            height=400 + len(df) * 30 + 50 # 增加图表高度以适应文本框
+        )
+
+        fig.show()
+        pass
+
+    async def _evals(self,prompt_id, OutputFormat, ExtraFormats_list = [],**kwargs):
+        
+        status,score, total, bad_case = await self.intellect_format_eval(
+            prompt_id=prompt_id,
+            OutputFormat = OutputFormat,
+            ExtraFormats = ExtraFormats_list,
+            version = None,
+            **kwargs
+        )
+        self.df.loc[len(self.df)] = {"name":prompt_id,
+                           'status':status,"score":score,
+                           "total":total,"bad_case":bad_case}
+
+# 整体测试d, 测试未通过d, 大模型调整再测试, 依旧不通过, 大模型裂变, 仍不通过, 互换人力