better：优化基于V3的心流效果，优化温度和prompt

2025-04-09 23:24:09 +08:00
parent f3d6e7cfa5
commit 451d0c9a32
6 changed files with 48 additions and 91 deletions
--- a/src/heart_flow/sub_heartflow.py
+++ b/src/heart_flow/sub_heartflow.py
@@ -13,6 +13,9 @@ from src.common.database import db
 from typing import Union
 from src.individuality.individuality import Individuality
 import random
+from src.plugins.chat.chat_stream import ChatStream
+from src.plugins.person_info.relationship_manager import relationship_manager
+from src.plugins.chat.utils import get_recent_group_speaker

 subheartflow_config = LogConfig(
    # 使用海马体专用样式
@@ -42,7 +45,7 @@ class SubHeartflow:
        self.past_mind = []
        self.current_state: CurrentState = CurrentState()
        self.llm_model = LLM_request(
-            model=global_config.llm_sub_heartflow, temperature=0.5, max_tokens=600, request_type="sub_heart_flow"
+            model=global_config.llm_sub_heartflow, temperature=0.3, max_tokens=600, request_type="sub_heart_flow"
        )

        self.main_heartflow_info = ""
@@ -58,6 +61,8 @@ class SubHeartflow:
        self.observations: list[Observation] = []

        self.running_knowledges = []
+        
+        self.bot_name = global_config.BOT_NICKNAME

    def add_observation(self, observation: Observation):
        """添加一个新的observation对象到列表中，如果已存在相同id的observation则不添加"""
@@ -106,56 +111,11 @@ class SubHeartflow:
            ):  # 5分钟无回复/不在场，销毁
                logger.info(f"子心流 {self.subheartflow_id} 已经5分钟没有激活，正在销毁...")
                break  # 退出循环，销毁自己
-
-    # async def do_a_thinking(self):
-    #     current_thinking_info = self.current_mind
-    #     mood_info = self.current_state.mood
-
-    #     observation = self.observations[0]
-    #     chat_observe_info = observation.observe_info
-    #     # print(f"chat_observe_info：{chat_observe_info}")
-
-    #     # 调取记忆
-    #     related_memory = await HippocampusManager.get_instance().get_memory_from_text(
-    #         text=chat_observe_info, max_memory_num=2, max_memory_length=2, max_depth=3, fast_retrieval=False
-    #     )
-
-    #     if related_memory:
-    #         related_memory_info = ""
-    #         for memory in related_memory:
-    #             related_memory_info += memory[1]
-    #     else:
-    #         related_memory_info = ""
-
-    #     # print(f"相关记忆：{related_memory_info}")
-
-    #     schedule_info = bot_schedule.get_current_num_task(num=1, time_info=False)
-
-    #     prompt = ""
-    #     prompt += f"你刚刚在做的事情是：{schedule_info}\n"
-    #     # prompt += f"麦麦的总体想法是：{self.main_heartflow_info}\n\n"
-    #     prompt += f"你{self.personality_info}\n"
-    #     if related_memory_info:
-    #         prompt += f"你想起来你之前见过的回忆：{related_memory_info}。\n以上是你的回忆，不一定是目前聊天里的人说的，也不一定是现在发生的事情，请记住。\n"
-    #     prompt += f"刚刚你的想法是{current_thinking_info}。\n"
-    #     prompt += "-----------------------------------\n"
-    #     prompt += f"现在你正在上网，和qq群里的网友们聊天，群里正在聊的话题是：{chat_observe_info}\n"
-    #     prompt += f"你现在{mood_info}\n"
-    #     prompt += "现在你接下去继续思考，产生新的想法，不要分点输出，输出连贯的内心独白，不要太长，"
-    #     prompt += "但是记得结合上述的消息，要记得维持住你的人设，关注聊天和新内容，不要思考太多:"
-    #     response, reasoning_content = await self.llm_model.generate_response_async(prompt)
-
-    #     self.update_current_mind(response)
-
-    #     self.current_mind = response
-    #     logger.debug(f"prompt:\n{prompt}\n")
-    #     logger.info(f"麦麦的脑内状态：{self.current_mind}")
-
    async def do_observe(self):
        observation = self.observations[0]
        await observation.observe()

-    async def do_thinking_before_reply(self, message_txt):
+    async def do_thinking_before_reply(self, message_txt:str, sender_name:str, chat_stream:ChatStream):
        current_thinking_info = self.current_mind
        mood_info = self.current_state.mood
        # mood_info = "你很生气，很愤怒"
@@ -164,7 +124,7 @@ class SubHeartflow:
        # print(f"chat_observe_info：{chat_observe_info}")

        # 开始构建prompt
-        prompt_personality = "你"
+        prompt_personality = f"你的名字是{self.bot_name},你"
        # person
        individuality = Individuality.get_instance()

@@ -178,6 +138,25 @@ class SubHeartflow:
        identity_detail = individuality.identity.identity_detail
        random.shuffle(identity_detail)
        prompt_personality += f",{identity_detail[0]}"
+        
+        # 关系
+        who_chat_in_group = [
+            (chat_stream.user_info.platform, chat_stream.user_info.user_id, chat_stream.user_info.user_nickname)
+        ]
+        who_chat_in_group += get_recent_group_speaker(
+            chat_stream.stream_id,
+            (chat_stream.user_info.platform, chat_stream.user_info.user_id),
+            limit=global_config.MAX_CONTEXT_SIZE,
+        )
+
+        relation_prompt = ""
+        for person in who_chat_in_group:
+            relation_prompt += await relationship_manager.build_relationship_info(person)
+
+        relation_prompt_all = (
+            f"{relation_prompt}关系等级越大，关系越好，请分析聊天记录，"
+            f"根据你和说话者{sender_name}的关系和态度进行回复，明确你的立场和情感。"
+        )

        # 调取记忆
        related_memory = await HippocampusManager.get_instance().get_memory_from_text(
@@ -204,6 +183,7 @@ class SubHeartflow:

        prompt = ""
        # prompt += f"麦麦的总体想法是：{self.main_heartflow_info}\n\n"
+        prompt += f"{relation_prompt_all}\n"
        prompt += f"{prompt_personality}\n"
        prompt += f"你刚刚在做的事情是：{schedule_info}\n"
        if related_memory_info:
@@ -214,9 +194,10 @@ class SubHeartflow:
        prompt += "-----------------------------------\n"
        prompt += f"现在你正在上网，和qq群里的网友们聊天，群里正在聊的话题是：{chat_observe_info}\n"
        prompt += f"你现在{mood_info}\n"
-        prompt += f"你注意到有人刚刚说：{message_txt}\n"
-        prompt += "现在你接下去继续思考，产生新的想法，不要分点输出，输出连贯的内心独白，不要太长，"
-        prompt += "记得结合上述的消息，要记得维持住你的人设，注意自己的名字，关注有人刚刚说的内容，不要思考太多:"
+        prompt += f"你注意到{sender_name}刚刚说：{message_txt}\n"
+        prompt += "现在你接下去继续浅浅思考，产生新的想法，不要分点输出，输出连贯的内心独白，不要太长，"
+        prompt += "思考时可以想想如何对群聊内容进行回复。请注意不要输出多余内容(包括前后缀，冒号和引号，括号，表情等)，"
+        prompt += f"记得结合上述的消息，要记得维持住你的人设，注意你就是{self.bot_name}，{self.bot_name}指的就是你。"

        try:
            response, reasoning_content = await self.llm_model.generate_response_async(prompt)
@@ -235,7 +216,7 @@ class SubHeartflow:
        # print("麦麦回复之后脑袋转起来了")

        # 开始构建prompt
-        prompt_personality = "你"
+        prompt_personality = f"你的名字是{self.bot_name},你"
        # person
        individuality = Individuality.get_instance()

--- a/src/plugins/chat_module/think_flow_chat/think_flow_chat.py
+++ b/src/plugins/chat_module/think_flow_chat/think_flow_chat.py
@@ -298,7 +298,9 @@ class ThinkFlowChat:
                try:
                    timer1 = time.time()
                    current_mind,past_mind = await heartflow.get_subheartflow(chat.stream_id).do_thinking_before_reply(
-                        message.processed_plain_text
+                        message_txt = message.processed_plain_text,
+                        sender_name = message.message_info.user_info.user_nickname,
+                        chat_stream = chat
                    )
                    timer2 = time.time()
                    timing_results["思考前脑内状态"] = timer2 - timer1
--- a/src/plugins/chat_module/think_flow_chat/think_flow_generator.py
+++ b/src/plugins/chat_module/think_flow_chat/think_flow_generator.py
@@ -25,7 +25,7 @@ logger = get_module_logger("llm_generator", config=llm_config)
 class ResponseGenerator:
    def __init__(self):
        self.model_normal = LLM_request(
-            model=global_config.llm_normal, temperature=0.8, max_tokens=256, request_type="response_heartflow"
+            model=global_config.llm_normal, temperature=0.6, max_tokens=256, request_type="response_heartflow"
        )

        self.model_sum = LLM_request(
--- a/src/plugins/chat_module/think_flow_chat/think_flow_prompt_builder.py
+++ b/src/plugins/chat_module/think_flow_chat/think_flow_prompt_builder.py
@@ -26,30 +26,7 @@ class PromptBuilder:
        individuality = Individuality.get_instance()
        prompt_personality = individuality.get_prompt(type="personality", x_person=2, level=1)
        prompt_identity = individuality.get_prompt(type="identity", x_person=2, level=1)
-        # 关系
-        who_chat_in_group = [
-            (chat_stream.user_info.platform, chat_stream.user_info.user_id, chat_stream.user_info.user_nickname)
-        ]
-        who_chat_in_group += get_recent_group_speaker(
-            stream_id,
-            (chat_stream.user_info.platform, chat_stream.user_info.user_id),
-            limit=global_config.MAX_CONTEXT_SIZE,
-        )

-        relation_prompt = ""
-        for person in who_chat_in_group:
-            relation_prompt += await relationship_manager.build_relationship_info(person)
-
-        relation_prompt_all = (
-            f"{relation_prompt}关系等级越大，关系越好，请分析聊天记录，"
-            f"根据你和说话者{sender_name}的关系和态度进行回复，明确你的立场和情感。"
-        )
-
-        # 心情
-        mood_manager = MoodManager.get_instance()
-        mood_prompt = mood_manager.get_prompt()
-
-        logger.info(f"心情prompt: {mood_prompt}")

        # 日程构建
        # schedule_prompt = f'''你现在正在做的事情是：{bot_schedule.get_current_num_task(num = 1,time_info = False)}'''
@@ -101,18 +78,16 @@ class PromptBuilder:
        logger.info("开始构建prompt")

        prompt = f"""
-        {relation_prompt_all}\n
 {chat_target}
 {chat_talking_prompt}
-你刚刚脑子里在想：
-{current_mind_info}
 现在"{sender_name}"说的:{message_txt}。引起了你的注意，你想要在群里发言发言或者回复这条消息。\n
 你的网名叫{global_config.BOT_NICKNAME}，有人也叫你{"/".join(global_config.BOT_ALIAS_NAMES)}，{prompt_personality} {prompt_identity}。
 你正在{chat_target_2},现在请你读读之前的聊天记录，然后给出日常且口语化的回复，平淡一些，
-尽量简短一些。{keywords_reaction_prompt}请注意把握聊天内容，不要回复的太有条理，可以有个性。{prompt_ger}
-请回复的平淡一些，简短一些，说中文，不要刻意突出自身学科背景，尽量不要说你说过的话 
-请注意不要输出多余内容(包括前后缀，冒号和引号，括号，表情等)，只输出回复内容。
-{moderation_prompt}不要输出多余内容(包括前后缀，冒号和引号，括号，表情包，at或 @等 )。"""
+你刚刚脑子里在想：
+{current_mind_info}
+回复尽量简短一些。{keywords_reaction_prompt}请注意把握聊天内容，不要回复的太有条理，可以有个性。{prompt_ger}
+请回复的平淡一些，简短一些，说中文，不要刻意突出自身学科背景，尽量不要说你说过的话 ，注意只输出回复内容。
+{moderation_prompt}。注意：不要输出多余内容(包括前后缀，冒号和引号，括号，表情包，at或 @等 )。"""

        return prompt

--- a/src/plugins/schedule/schedule_generator.py
+++ b/src/plugins/schedule/schedule_generator.py
@@ -32,7 +32,7 @@ class ScheduleGenerator:
        # 使用离线LLM模型
        self.llm_scheduler_all = LLM_request(
            model=global_config.llm_reasoning,
-            temperature=global_config.SCHEDULE_TEMPERATURE,
+            temperature=global_config.SCHEDULE_TEMPERATURE+0.3,
            max_tokens=7000,
            request_type="schedule",
        )
--- a/template/bot_config_template.toml
+++ b/template/bot_config_template.toml
@@ -60,7 +60,7 @@ appearance = "用几句话描述外貌特征" # 外貌特征
 enable_schedule_gen = true # 是否启用日程表(尚未完成)
 prompt_schedule_gen = "用几句话描述描述性格特点或行动规律，这个特征会用来生成日程表"
 schedule_doing_update_interval = 900 # 日程表更新间隔 单位秒
-schedule_temperature = 0.3 # 日程表温度，建议0.3-0.6
+schedule_temperature = 0.2 # 日程表温度，建议0.2-0.5
 time_zone = "Asia/Shanghai" # 给你的机器人设置时区，可以解决运行电脑时区和国内时区不同的情况，或者模拟国外留学生日程

 [platforms] # 必填项目，填写每个平台适配器提供的链接
@@ -239,12 +239,11 @@ provider = "SILICONFLOW"
 pri_in = 0
 pri_out = 0

-[model.llm_sub_heartflow] #心流：建议使用qwen2.5 7b
-# name = "Pro/Qwen/Qwen2.5-7B-Instruct"
-name = "Qwen/Qwen2.5-32B-Instruct"
+[model.llm_sub_heartflow] #子心流：建议使用V3级别
+name = "Pro/deepseek-ai/DeepSeek-V3"
 provider = "SILICONFLOW"
-pri_in = 1.26
-pri_out = 1.26
+pri_in = 2
+pri_out = 8

 [model.llm_heartflow] #心流：建议使用qwen2.5 32b
 # name = "Pro/Qwen/Qwen2.5-7B-Instruct"