161 lines
6.5 KiB
Python
161 lines
6.5 KiB
Python
"""
|
||
llmcheck 模式：
|
||
此模式的一些参数不会在配置文件中显示，要修改请在可变参数下修改
|
||
此模式的特点：
|
||
1.在群聊内的连续对话场景下，使用大语言模型来判断回复概率
|
||
2.非连续对话场景,使用mxp模式的意愿管理器(可另外配置)
|
||
3.默认配置的是model_v3,当前参数适用于deepseek-v3-0324
|
||
|
||
继承自其他模式,实质上仅重写get_reply_probability方法,未来可能重构成一个插件,可方便地组装到其他意愿模式上。
|
||
目前的使用方式是拓展到其他意愿管理模式
|
||
|
||
"""
|
||
import time
|
||
from loguru import logger
|
||
from ..models.utils_model import LLM_request
|
||
from ...config.config import global_config
|
||
from ..chat.chat_stream import ChatStream
|
||
from ..chat.utils import get_recent_group_detailed_plain_text
|
||
from .willing_manager import BaseWillingManager
|
||
from .mode_mxp import MxpWillingManager
|
||
import re
|
||
from functools import wraps
|
||
|
||
|
||
def is_continuous_chat(self, message_id: str):
|
||
# 判断是否是连续对话，出于成本考虑，默认限制5条
|
||
willing_info = self.ongoing_messages[message_id]
|
||
chat_id = willing_info.chat_id
|
||
group_info = willing_info.group_info
|
||
config = self.global_config
|
||
length = 5
|
||
if chat_id:
|
||
chat_talking_text = get_recent_group_detailed_plain_text(
|
||
chat_id, limit=length, combine=True
|
||
)
|
||
if group_info:
|
||
if str(config.BOT_QQ) in chat_talking_text:
|
||
return True
|
||
else:
|
||
return False
|
||
return False
|
||
|
||
def llmcheck_decorator(trigger_condition_func):
|
||
def decorator(func):
|
||
@wraps(func)
|
||
def wrapper(self, message_id: str):
|
||
if trigger_condition_func(self, message_id):
|
||
# 满足条件，走llm流程
|
||
return self.get_llmreply_probability(message_id)
|
||
else:
|
||
# 不满足条件，走默认流程
|
||
return func(self, message_id)
|
||
return wrapper
|
||
return decorator
|
||
|
||
|
||
class LlmcheckWillingManager(MxpWillingManager):
|
||
|
||
def __init__(self):
|
||
super().__init__()
|
||
self.model_v3 = LLM_request(model=global_config.llm_normal, temperature=0.3)
|
||
|
||
|
||
|
||
async def get_llmreply_probability(self, message_id: str):
|
||
message_info = self.ongoing_messages[message_id]
|
||
chat_id = message_info.chat_id
|
||
config = self.global_config
|
||
# 获取信息的长度
|
||
length = 5
|
||
if message_info.group_info and config:
|
||
if message_info.group_info.group_id not in config.talk_allowed_groups:
|
||
reply_probability = 0
|
||
return reply_probability
|
||
|
||
current_date = time.strftime("%Y-%m-%d", time.localtime())
|
||
current_time = time.strftime("%H:%M:%S", time.localtime())
|
||
chat_talking_prompt = ""
|
||
if chat_id:
|
||
chat_talking_prompt = get_recent_group_detailed_plain_text(
|
||
chat_id, limit=length, combine=True
|
||
)
|
||
else:
|
||
return 0
|
||
|
||
# if is_mentioned_bot:
|
||
# return 1.0
|
||
prompt = f"""
|
||
假设你正在查看一个群聊，你在这个群聊里的网名叫{global_config.BOT_NICKNAME}，你还有很多别名: {"/".join(global_config.BOT_ALIAS_NAMES)}，
|
||
现在群里聊天的内容是{chat_talking_prompt}，
|
||
今天是{current_date}，现在是{current_time}。
|
||
综合群内的氛围和你自己之前的发言，给出你认为**最新的消息**需要你回复的概率，数值在0到1之间。请注意，群聊内容杂乱，很多时候对话连续，但很可能不是在和你说话。
|
||
如果最新的消息和你之前的发言在内容上连续，或者提到了你的名字或者称谓，将其视作明确指向你的互动，给出高于0.8的概率。如果现在是睡眠时间，直接概率为0。如果话题内容与你之前不是紧密相关，请不要给出高于0.1的概率。
|
||
请注意是判断概率，而不是编写回复内容，
|
||
仅输出在0到1区间内的概率值，不要给出你的判断依据。
|
||
"""
|
||
|
||
content_check, reasoning_check, _ = await self.model_v3.generate_response(prompt)
|
||
# logger.info(f"{prompt}")
|
||
logger.info(f"{content_check} {reasoning_check}")
|
||
probability = self.extract_marked_probability(content_check)
|
||
# 兴趣系数修正 无关激活效率太高，暂时停用，待新记忆系统上线后调整
|
||
probability += (message_info.interested_rate * 0.25)
|
||
probability = min(1.0, probability)
|
||
if probability <= 0.1:
|
||
probability = min(0.03, probability)
|
||
if probability >= 0.8:
|
||
probability = max(probability, 0.90)
|
||
|
||
# 当前表情包理解能力较差，少说就少错
|
||
if message_info.is_emoji:
|
||
probability *= global_config.emoji_response_penalty
|
||
|
||
return probability
|
||
|
||
@staticmethod
|
||
def extract_marked_probability(text):
|
||
"""提取带标记的概率值 该方法主要用于测试微调prompt阶段"""
|
||
text = text.strip()
|
||
pattern = r'##PROBABILITY_START##(.*?)##PROBABILITY_END##'
|
||
match = re.search(pattern, text, re.DOTALL)
|
||
if match:
|
||
prob_str = match.group(1).strip()
|
||
# 处理百分比（65% → 0.65）
|
||
if '%' in prob_str:
|
||
return float(prob_str.replace('%', '')) / 100
|
||
# 处理分数（2/3 → 0.666...）
|
||
elif '/' in prob_str:
|
||
numerator, denominator = map(float, prob_str.split('/'))
|
||
return numerator / denominator
|
||
# 直接处理小数
|
||
else:
|
||
return float(prob_str)
|
||
|
||
percent_match = re.search(r'(\d{1,3})%', text) # 65%
|
||
decimal_match = re.search(r'(0\.\d+|1\.0+)', text) # 0.65
|
||
fraction_match = re.search(r'(\d+)/(\d+)', text) # 2/3
|
||
try:
|
||
if percent_match:
|
||
prob = float(percent_match.group(1)) / 100
|
||
elif decimal_match:
|
||
prob = float(decimal_match.group(0))
|
||
elif fraction_match:
|
||
numerator, denominator = map(float, fraction_match.groups())
|
||
prob = numerator / denominator
|
||
else:
|
||
return 0 # 无匹配格式
|
||
|
||
# 验证范围是否合法
|
||
if 0 <= prob <= 1:
|
||
return prob
|
||
return 0
|
||
except (ValueError, ZeroDivisionError):
|
||
return 0
|
||
|
||
@llmcheck_decorator(is_continuous_chat)
|
||
def get_reply_probability(self, message_id):
|
||
return super().get_reply_probability(
|
||
message_id
|
||
)
|