gardel/Mofox-Core
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat:合并工具调用模型和心流模型

Browse Source
This commit is contained in:
SengokuCola
2025-04-24 14:18:41 +08:00
parent 2871e4efc2
commit f8450f705a
22 changed files with 973 additions and 331 deletions
Download Patch File Download Diff File Expand all files Collapse all files

194
src/heart_flow/sub_heartflow.py
View File

@@ -18,10 +18,9 @@ from src.plugins.chat.chat_stream import chat_manager
import math
from src.plugins.heartFC_chat.heartFC_chat import HeartFChatting
from src.plugins.heartFC_chat.normal_chat import NormalChat
# from src.do_tool.tool_use import ToolUser
from src.do_tool.tool_use import ToolUser
from src.heart_flow.mai_state_manager import MaiStateInfo
from src.plugins.utils.json_utils import safe_json_dumps, process_llm_tool_response, normalize_llm_response, process_llm_tool_calls
# 定义常量 (从 interest.py 移动过来)
MAX_INTEREST = 15.0
@@ -54,8 +53,9 @@ def init_prompt():
# prompt += "你注意到{sender_name}刚刚说：{message_txt}\n"
prompt += "现在请你根据刚刚的想法继续思考，思考时可以想想如何对群聊内容进行回复，要不要对群里的话题进行回复，关注新话题，可以适当转换话题，大家正在说的话才是聊天的主题。\n"
prompt += "回复的要求是：平淡一些，简短一些，说中文，如果你要回复，最好只回复一个人的一个话题\n"
prompt += "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)，不要带有括号和动作描写。不要回复自己的发言，尽量不要说你说过的话。"
prompt += "现在请你{hf_do_next}，不要分点输出,生成内心想法，文字不要浮夸"
prompt += "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)，不要带有括号和动作描写。不要回复自己的发言，尽量不要说你说过的话。\n"
prompt += "现在请你先{hf_do_next}，不要分点输出,生成内心想法，文字不要浮夸"
prompt += "在输出完想法后，请你思考应该使用什么工具。如果你需要做某件事，来对消息和你的回复进行处理，请使用工具。\n"
Prompt(prompt, "sub_heartflow_prompt_before")
@@ -114,6 +114,8 @@ class InterestChatting:
self.above_threshold = False
self.start_hfc_probability = 0.0
def add_interest_dict(self, message: MessageRecv, interest_value: float, is_mentioned: bool):
self.interest_dict[message.message_info.message_id] = (message, interest_value, is_mentioned)
@@ -291,6 +293,8 @@ class SubHeartflow:
)
self.log_prefix = chat_manager.get_stream_name(self.subheartflow_id) or self.subheartflow_id
self.structured_info = {}
async def add_time_current_state(self, add_time: float):
self.current_state_time += add_time
@@ -477,58 +481,63 @@ class SubHeartflow:
logger.info(f"{self.log_prefix} 子心流后台任务已停止。")
async def do_thinking_before_reply(
self,
extra_info: str,
obs_id: list[str] = None,
):
async def do_thinking_before_reply(self):
"""
在回复前进行思考，生成内心想法并收集工具调用结果
返回:
tuple: (current_mind, past_mind) 当前想法和过去的想法列表
"""
# 更新活跃时间
self.last_active_time = time.time()
# ---------- 1. 准备基础数据 ----------
# 获取现有想法和情绪状态
current_thinking_info = self.current_mind
mood_info = self.chat_state.mood
# 获取观察对象
observation = self._get_primary_observation()
chat_observe_info = ""
if obs_id:
try:
chat_observe_info = observation.get_observe_info(obs_id)
logger.debug(f"[{self.subheartflow_id}] Using specific observation IDs: {obs_id}")
except Exception as e:
logger.error(
f"[{self.subheartflow_id}] Error getting observe info with IDs {obs_id}: {e}. Falling back."
)
chat_observe_info = observation.get_observe_info()
else:
chat_observe_info = observation.get_observe_info()
# logger.debug(f"[{self.subheartflow_id}] Using default observation info.")
extra_info_prompt = ""
if extra_info:
for tool_name, tool_data in extra_info.items():
extra_info_prompt += f"{tool_name} 相关信息:\n"
for item in tool_data:
extra_info_prompt += f"- {item['name']}: {item['content']}\n"
else:
extra_info_prompt = "无工具信息。\n"
if not observation:
logger.error(f"[{self.subheartflow_id}] 无法获取观察对象")
self.update_current_mind("(我没看到任何聊天内容...)")
return self.current_mind, self.past_mind
# 获取观察内容
chat_observe_info = observation.get_observe_info()
# ---------- 2. 准备工具和个性化数据 ----------
# 初始化工具
tool_instance = ToolUser()
tools = tool_instance._define_tools()
# 获取个性化信息
individuality = Individuality.get_instance()
# 构建个性部分
prompt_personality = f"你的名字是{individuality.personality.bot_nickname}，你"
prompt_personality += individuality.personality.personality_core
# 随机添加个性侧面
if individuality.personality.personality_sides:
random_side = random.choice(individuality.personality.personality_sides)
prompt_personality += f"，{random_side}"
# 随机添加身份细节
if individuality.identity.identity_detail:
random_detail = random.choice(individuality.identity.identity_detail)
prompt_personality += f"，{random_detail}"
# 获取当前时间
time_now = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
# ---------- 3. 构建思考指导部分 ----------
# 创建本地随机数生成器，基于分钟数作为种子
local_random = random.Random()
current_minute = int(time.strftime("%M"))
local_random.seed(current_minute)
# 思考指导选项和权重
hf_options = [
("继续生成你在这个聊天中的想法，在原来想法的基础上继续思考", 0.7),
("生成你在这个聊天中的想法，在原来的想法上尝试新的话题", 0.1),
@@ -536,12 +545,17 @@ class SubHeartflow:
("继续生成你在这个聊天中的想法，进行深入思考", 0.1),
]
# 加权随机选择思考指导
hf_do_next = local_random.choices(
[option[0] for option in hf_options], weights=[option[1] for option in hf_options], k=1
[option[0] for option in hf_options],
weights=[option[1] for option in hf_options],
k=1
)[0]
# ---------- 4. 构建最终提示词 ----------
# 获取提示词模板并填充数据
prompt = (await global_prompt_manager.get_prompt_async("sub_heartflow_prompt_before")).format(
extra_info=extra_info_prompt,
extra_info="", # 可以在这里添加额外信息
prompt_personality=prompt_personality,
bot_name=individuality.personality.bot_nickname,
current_thinking_info=current_thinking_info,
@@ -551,26 +565,104 @@ class SubHeartflow:
hf_do_next=hf_do_next,
)
prompt = await relationship_manager.convert_all_person_sign_to_person_name(prompt)
prompt = parse_text_timestamps(prompt, mode="lite")
logger.debug(f"[{self.subheartflow_id}] 心流思考prompt:\n{prompt}\n")
logger.debug(f"[{self.subheartflow_id}] 心流思考提示词构建完成")
# ---------- 5. 执行LLM请求并处理响应 ----------
content = "" # 初始化内容变量
reasoning_content = "" # 初始化推理内容变量
try:
response, reasoning_content = await self.llm_model.generate_response_async(prompt)
logger.debug(f"[{self.subheartflow_id}] 心流思考结果:\n{response}\n")
if not response:
response = "(不知道该想些什么...)"
logger.warning(f"[{self.subheartflow_id}] LLM 返回空结果，思考失败。")
# 调用LLM生成响应
response = await self.llm_model.generate_response_tool_async(prompt=prompt, tools=tools)
# 标准化响应格式
success, normalized_response, error_msg = normalize_llm_response(
response, log_prefix=f"[{self.subheartflow_id}] "
)
if not success:
# 处理标准化失败情况
logger.warning(f"[{self.subheartflow_id}] {error_msg}")
content = "LLM响应格式无法处理"
else:
# 从标准化响应中提取内容
if len(normalized_response) >= 2:
content = normalized_response[0]
reasoning_content = normalized_response[1] if len(normalized_response) > 1 else ""
# 处理可能的工具调用
if len(normalized_response) == 3:
# 提取并验证工具调用
success, valid_tool_calls, error_msg = process_llm_tool_calls(
normalized_response, log_prefix=f"[{self.subheartflow_id}] "
)
if success and valid_tool_calls:
# 记录工具调用信息
tool_calls_str = ", ".join([
call.get("function", {}).get("name", "未知工具")
for call in valid_tool_calls
])
logger.info(f"[{self.subheartflow_id}] 模型请求调用{len(valid_tool_calls)}个工具: {tool_calls_str}")
# 收集工具执行结果
await self._execute_tool_calls(valid_tool_calls, tool_instance)
elif not success:
logger.warning(f"[{self.subheartflow_id}] {error_msg}")
except Exception as e:
logger.error(f"[{self.subheartflow_id}] 内心独白获取失败: {e}")
response = "(思考时发生错误...)"
# 处理总体异常
logger.error(f"[{self.subheartflow_id}] 执行LLM请求或处理响应时出错: {e}")
logger.error(traceback.format_exc())
content = "思考过程中出现错误"
self.update_current_mind(response)
# 记录最终思考结果
logger.debug(f"[{self.subheartflow_id}] 心流思考结果:\n{content}\n")
# 处理空响应情况
if not content:
content = "(不知道该想些什么...)"
logger.warning(f"[{self.subheartflow_id}] LLM返回空结果，思考失败。")
# ---------- 6. 更新思考状态并返回结果 ----------
# 更新当前思考内容
self.update_current_mind(content)
return self.current_mind, self.past_mind
async def _execute_tool_calls(self, tool_calls, tool_instance):
"""
执行一组工具调用并收集结果
参数:
tool_calls: 工具调用列表
tool_instance: 工具使用器实例
"""
tool_results = []
structured_info = {} # 动态生成键
# 执行所有工具调用
for tool_call in tool_calls:
try:
result = await tool_instance._execute_tool_call(tool_call)
if result:
tool_results.append(result)
# 使用工具名称作为键
tool_name = result["name"]
if tool_name not in structured_info:
structured_info[tool_name] = []
structured_info[tool_name].append({
"name": result["name"],
"content": result["content"]
})
except Exception as tool_e:
logger.error(f"[{self.subheartflow_id}] 工具执行失败: {tool_e}")
# 如果有工具结果，记录并更新结构化信息
if structured_info:
logger.debug(f"工具调用收集到结构化信息: {safe_json_dumps(structured_info, ensure_ascii=False)}")
self.structured_info = structured_info
def update_current_mind(self, response):
self.past_mind.append(self.current_mind)