123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283 |
- import json
- import datetime
- import random
- from tqdm import tqdm
- from openai import OpenAI
- from pymysql.cursors import DictCursor
- from pqai_agent.database import MySQLManager
- from pqai_agent.agents.message_push_agent import MessagePushAgent
- from pqai_agent.agents.message_reply_agent import MessageReplyAgent
- def fetch_deepseek_completion(prompt, output_type='text'):
- """
- deep_seek方法
- """
- client = OpenAI(
- api_key='sk-cfd2df92c8864ab999d66a615ee812c5',
- base_url="https://api.deepseek.com"
- )
- # get response format
- if output_type == "json":
- response_format = {"type": "json_object"}
- else:
- response_format = {"type": "text"}
- chat_completion = client.chat.completions.create(
- messages=[
- {
- "role": "user",
- "content": prompt,
- }
- ],
- model="deepseek-chat",
- response_format=response_format,
- )
- response = chat_completion.choices[0].message.content
- if output_type == "json":
- response_json = json.loads(response)
- return response_json
- return response
- def get_profile_info(user_id_, user_type):
- match user_type:
- case "user":
- sql = f"""
- select iconurl as 'avatar', profile_data_v1 as 'profile'
- from third_party_user where third_party_user_id = %s;
- """
- case "staff":
- sql = f"""
- select agent_profile as 'profile'
- from qywx_employee where third_party_user_id = %s;
- """
- case _:
- raise ValueError("user_type must be 'user' or 'staff'")
- return mysql_client.select(sql, cursor_type=DictCursor, args=(user_id_,))
- def evaluate_conversation_quality_task(dialogue_history, user_profile_, agent_profile):
- """
- :param dialogue_history:
- :param user_profile_:
- :param agent_profile:
- :return:
- """
- output_format = {
- "1.1": {
- "score": 5,
- "reason": ""
- },
- "1.2": {
- "score": 8,
- "reason": "reason"
- },
- "1.3": {
- "score": 10,
- "reason": "reason"
- },
- "1.4": {
- "score": 10,
- "reason": "reason"
- },
- "1.5": {
- "score": 10,
- "reason": "reason"
- },
- "1.6": {
- "score": 10,
- "reason": "reason"
- },
- "2.1": {
- "score": 9,
- "reason": "reason"
- },
- "2.2": {
- "score": 10,
- "reason": "reason"
- },
- "2.3": {
- "score": 10,
- "reason": "reason"
- },
- "total_score": "total_score",
- "improvement_suggestions": "suggestions",
- }
- prompt_ = f"""
- 你是一名优秀的 agent 评估员,请根据以下场景和输入,对该 agent 的回复能力进行评估,用分数量化
- 场景:
- 智能体对话场景, 智能体(agent)和用户(user)进行对话聊天
- 输入:
- agent 的人设:agent_profile: {agent_profile}
- 用户的人设: user_profile: {user_profile_}
- 对话历史:dialogue_history: {dialogue_history}
- 评估标准, 满分为 100分,拆分到以下每一个小项,每一个小项的得分表示该小项的能力,60% 的分表示及格,80% 的分表示优秀:
- 1. 对话能力(30分)
- 1.1 语言是否流畅(10分)
- 1.2 上下文是否连贯,语义是否一致(10分)
- 1.3 agent 是否感知用户结束聊天的意图并且适当结束聊天(10分)
- 1.4 agent 回复消息的时间间隔是否合理,符合真人对话规律 (10分)
- 1.5 agent 回复的消息是否具有高情商,互动能力是否好,能否和用户共情,提升用户的情感体验 (20分)
- 1.6 agent 回复的消息是否解决了用户提出的问题 (10分)
- 2. 角色一致性(30分)
- 2.1 agent 语言风格是否符合agent人设(10分)
- 2.2 agent 语言风格是否适合用户人设(10分)
- 2.3 agent 回复内容不要超越用户的认知上限(10分)
- 输出:
- 输出为 json 格式,输出格式规范 {output_format}
- """
- return prompt_
- def evaluate_push_agent_prompt(dialogue_history, push_message, user_profile_, agent_profile):
- """
- :param dialogue_history:
- :param push_message:
- :param user_profile_:
- :param agent_profile:
- :return:
- """
- output_format = {
- "1.1": {
- "score": 5,
- "reason": "push_message尝试联系用户的头像,但用户兴趣未明确提及戏曲"
- },
- "1.2": {
- "score": 8,
- "reason": "语言风格轻松友好,适合大多数用户,但未完全匹配用户特定风格"
- },
- "1.3": {
- "score": 10,
- "reason": "信息未超出用户认知范围"
- },
- "2.1": {
- "score": 9,
- "reason": "语言风格符合agent人设,友好且亲切"
- },
- "2.2": {
- "score": 10,
- "reason": "信息未超出agent人设的认知范围"
- },
- "3.1": {
- "score": 15,
- "reason": "push_message有潜力勾起用户兴趣,但未直接关联用户已知兴趣"
- },
- "3.2": {
- "score": 10,
- "reason": "信息真实"
- },
- "3.3": {
- "score": 12,
- "reason": "表现出一定的拟人化和情商,但共情程度可进一步提升"
- },
- "total_score": 79,
- "improvement_suggestions": "建议更深入地挖掘和利用用户已知的兴趣爱好来定制push_message,以增强相关性和用户参与度。同时,可以尝试更多共情的表达方式,以提升用户的情感体验。"
- }
- prompt_ = f"""
- 你是一名优秀的 agent 评估员,请根据以下场景和输入,对该 agent 的能力进行评估,用分数量化
- 场景:
- 智能体对话场景, 智能体(agent)向用户发起对话
- agent 需要通过分析 user 和 agent 直接的历史对话,以及 user 和 agent 的人设信息,向用户发送一条消息(push_message)
- 输入:
- agent 的人设:agent_profile: {agent_profile}
- 用户的人设: user_profile: {user_profile_}
- 对话历史:dialogue_history: {dialogue_history}
- agent 的唤起对话:push_message: {push_message}
- 评估标准, 满分为 100分,拆分到以下每一个小项,每一个小项的得分表示该小项的能力,60% 的分表示及格,80% 的分表示优秀:
- 1. push_message 的内容 和 user_profile的相关性(30分)
- 1.1 push_message 是否迎合用户的兴趣爱好 (满分 10分)
- 1.2 push_message 的语言风格是否适合用户语言风格 (满分 10分)
- 1.3 push_message 的信息是否超出用户的认知范围 (满分 10分)
- 2. push_message 和 agent_profile 的相关性(20分)
- 2.1 push_message 的语言风格是否符合 agent 人设(满分 10分)
- 2.2 push_message 的信息是否超出 agent人设的认知范围(满分 10分)
- 3. push_message 质量量化 (50分)
- 3.1 push_message 是否能勾起用户的兴趣,驱动用户聊天激情 (满分 25分)
- 3.2 push_message 的信息是否真实 (满分 10分)
- 3.3 push_message 是否具有拟人化,高情商,与用户共情,提升用户的情感体验(满分 15分)
- 输出:
- 输出为 json 格式,输出格式规范 {output_format}
- """
- return prompt_
- def evaluate_reply_agent(dialogue_history, reply_message, user_profile_, agent_profile):
- """
- :param dialogue_history:
- :param reply_message:
- :param user_profile_:
- :param agent_profile:
- :return:
- """
- return
- config = {
- 'host': 'rm-bp13g3ra2f59q49xs.mysql.rds.aliyuncs.com',
- 'port': 3306,
- 'user': 'wqsd',
- 'password': 'wqsd@2025',
- 'database': 'ai_agent',
- 'charset': 'utf8mb4'
- }
- mysql_client = MySQLManager(config)
- if __name__ == '__main__':
- import pqai_agent.logging_service
- pqai_agent.logging_service.setup_root_logger()
- with open("scripts/dialogue_data_set.json", "r", encoding="utf-8") as f:
- data = json.load(f)
- dialogues = random.sample(data, 100)
- F = []
- for sub_dialogues in tqdm(dialogues):
- try:
- # user 相关
- user_id = sub_dialogues['user_id']
- user_profile_response = get_profile_info(user_id, "user")
- user_profile, avatar = json.loads(user_profile_response[0]['profile']), user_profile_response[0]['avatar']
- user_profile['avatar'] = avatar
- user_profile['current_datetime'] = datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")
- # staff 相关
- staff_id = sub_dialogues['staff_id']
- staff_profile_response = get_profile_info(staff_id, "staff")
- staff_profile = json.loads(staff_profile_response[0]['profile'])
- user_profile['formatted_staff_profile'] = staff_profile
- push_agent = MessagePushAgent()
- # reply_agent = MessageReplyAgent()
- # message 相关
- message = sub_dialogues['dialogue_history']
- agent_message = push_agent.generate_message(
- context=user_profile,
- dialogue_history=message
- )
- prompt = evaluate_push_agent_prompt(message, agent_message, user_profile, staff_profile)
- # prompt = evaluate_conversation_quality_task(message, user_profile, staff_profile)
- response = fetch_deepseek_completion(prompt, output_type='json')
- obj = {
- "user_profile": user_profile,
- "agent_profile": staff_profile,
- "dialogue_history": message,
- "push_message": agent_message,
- "push_time": datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
- "evaluation_result": response
- }
- F.append(obj)
- except:
- continue
- with open("push_message_evaluation_result.json", "w", encoding="utf-8") as f:
- f.write(json.dumps(F, indent=4, ensure_ascii=False))
|