dev：修复记忆构建文本名称问题
This commit is contained in:
SengokuCola
@@ -1,7 +1,6 @@
|
||||
import asyncio
|
||||
import time
|
||||
import traceback
|
||||
import random # <-- 添加导入
|
||||
from typing import List, Optional, Dict, Any, Deque, Callable, Coroutine
|
||||
from collections import deque
|
||||
from src.plugins.chat.message import MessageRecv, BaseMessageInfo, MessageThinking, MessageSending
|
||||
@@ -14,17 +13,20 @@ from src.plugins.models.utils_model import LLMRequest
|
||||
from src.config.config import global_config
|
||||
from src.plugins.chat.utils_image import image_path_to_base64 # Local import needed after move
|
||||
from src.plugins.utils.timer_calculator import Timer # <--- Import Timer
|
||||
from src.plugins.heartFC_chat.heartFC_generator import HeartFCGenerator
|
||||
from src.do_tool.tool_use import ToolUser
|
||||
from src.plugins.emoji_system.emoji_manager import emoji_manager
|
||||
from src.plugins.utils.json_utils import process_llm_tool_calls, extract_tool_call_arguments
|
||||
from src.heart_flow.sub_mind import SubMind
|
||||
from src.heart_flow.observation import Observation
|
||||
from src.plugins.heartFC_chat.heartflow_prompt_builder import global_prompt_manager
|
||||
from src.plugins.heartFC_chat.heartflow_prompt_builder import global_prompt_manager, prompt_builder
|
||||
import contextlib
|
||||
from src.plugins.utils.chat_message_builder import num_new_messages_since
|
||||
from src.plugins.heartFC_chat.heartFC_Cycleinfo import CycleInfo
|
||||
from .heartFC_sender import HeartFCSender
|
||||
from src.plugins.chat.utils import process_llm_response
|
||||
from src.plugins.respon_info_catcher.info_catcher import info_catcher_manager
|
||||
from src.plugins.moods.moods import MoodManager
|
||||
from src.individuality.individuality import Individuality
|
||||
|
||||
|
||||
INITIAL_DURATION = 60.0
|
||||
@@ -181,12 +183,18 @@ class HeartFChatting:
|
||||
self.action_manager = ActionManager()
|
||||
|
||||
# 初始化状态控制
|
||||
self._initialized = False # 是否已初始化标志
|
||||
self._processing_lock = asyncio.Lock() # 处理锁(确保单次Plan-Replier-Sender周期)
|
||||
self._initialized = False
|
||||
self._processing_lock = asyncio.Lock()
|
||||
|
||||
# 依赖注入存储
|
||||
self.gpt_instance = HeartFCGenerator() # 文本回复生成器
|
||||
self.tool_user = ToolUser() # 工具使用实例
|
||||
# --- 移除 gpt_instance, 直接初始化 LLM 模型 ---
|
||||
# self.gpt_instance = HeartFCGenerator() # <-- 移除
|
||||
self.model_normal = LLMRequest( # <-- 新增 LLM 初始化
|
||||
model=global_config.llm_normal,
|
||||
temperature=global_config.llm_normal["temp"],
|
||||
max_tokens=256,
|
||||
request_type="response_heartflow",
|
||||
)
|
||||
self.tool_user = ToolUser()
|
||||
self.heart_fc_sender = HeartFCSender()
|
||||
|
||||
# LLM规划器配置
|
||||
@@ -401,16 +409,15 @@ class HeartFChatting:
|
||||
with Timer("决策", cycle_timers):
|
||||
planner_result = await self._planner(current_mind, cycle_timers)
|
||||
|
||||
|
||||
# 效果不太好，还没处理replan导致观察时间点改变的问题
|
||||
|
||||
|
||||
# action = planner_result.get("action", "error")
|
||||
# reasoning = planner_result.get("reasoning", "未提供理由")
|
||||
|
||||
# self._current_cycle.set_action_info(action, reasoning, False)
|
||||
|
||||
# 在获取规划结果后检查新消息
|
||||
|
||||
|
||||
# if await self._check_new_messages(planner_start_db_time):
|
||||
# if random.random() < 0.2:
|
||||
# logger.info(f"{self.log_prefix} 看到了新消息，麦麦决定重新观察和规划...")
|
||||
@@ -742,8 +749,8 @@ class HeartFChatting:
|
||||
# --- 使用 LLM 进行决策 --- #
|
||||
reasoning = "默认决策或获取决策失败"
|
||||
llm_error = False # LLM错误标志
|
||||
arguments = None # 初始化参数变量
|
||||
emoji_query = "" # <--- 在这里初始化 emoji_query
|
||||
arguments = None # 初始化参数变量
|
||||
emoji_query = "" # <--- 在这里初始化 emoji_query
|
||||
|
||||
try:
|
||||
# --- 构建提示词 ---
|
||||
@@ -756,7 +763,7 @@ class HeartFChatting:
|
||||
observed_messages_str, current_mind, self.sub_mind.structured_info, replan_prompt_str
|
||||
)
|
||||
|
||||
# --- 调用 LLM ---
|
||||
# --- 调用 LLM ---
|
||||
try:
|
||||
planner_tools = self.action_manager.get_planner_tool_definition()
|
||||
_response_text, _reasoning_content, tool_calls = await self.planner_llm.generate_response_tool_async(
|
||||
@@ -794,7 +801,7 @@ class HeartFChatting:
|
||||
first_tool_call = valid_tool_calls[0]
|
||||
tool_name = first_tool_call.get("function", {}).get("name")
|
||||
arguments = extract_tool_call_arguments(first_tool_call, None)
|
||||
|
||||
|
||||
# 3. 检查名称和参数
|
||||
expected_tool_name = "decide_reply_action"
|
||||
if tool_name == expected_tool_name and arguments is not None:
|
||||
@@ -808,13 +815,13 @@ class HeartFChatting:
|
||||
action = "no_reply"
|
||||
reasoning = f"LLM返回了未授权的动作: {extracted_action}"
|
||||
emoji_query = ""
|
||||
llm_error = False # 视为非LLM错误，只是逻辑修正
|
||||
llm_error = False # 视为非LLM错误，只是逻辑修正
|
||||
else:
|
||||
# 动作有效，使用提取的值
|
||||
action = extracted_action
|
||||
reasoning = arguments.get("reasoning", "未提供理由")
|
||||
emoji_query = arguments.get("emoji_query", "")
|
||||
llm_error = False # 成功处理
|
||||
llm_error = False # 成功处理
|
||||
# 记录决策结果
|
||||
logger.debug(
|
||||
f"{self.log_prefix}[要做什么]\nPrompt:\n{prompt}\n\n决策结果: {action}, 理由: {reasoning}, 表情查询: '{emoji_query}'"
|
||||
@@ -822,13 +829,13 @@ class HeartFChatting:
|
||||
elif tool_name != expected_tool_name:
|
||||
reasoning = f"LLM返回了非预期的工具: {tool_name}"
|
||||
logger.warning(f"{self.log_prefix}[Planner] {reasoning}")
|
||||
else: # arguments is None
|
||||
else: # arguments is None
|
||||
reasoning = f"无法提取工具 {tool_name} 的参数"
|
||||
logger.warning(f"{self.log_prefix}[Planner] {reasoning}")
|
||||
elif not success:
|
||||
reasoning = f"验证工具调用失败: {error_msg}"
|
||||
logger.warning(f"{self.log_prefix}[Planner] {reasoning}")
|
||||
else: # not valid_tool_calls
|
||||
else: # not valid_tool_calls
|
||||
reasoning = "LLM未返回有效的工具调用"
|
||||
logger.warning(f"{self.log_prefix}[Planner] {reasoning}")
|
||||
# 如果 llm_error 仍然是 True，说明在处理过程中有错误发生
|
||||
@@ -1058,9 +1065,13 @@ class HeartFChatting:
|
||||
# 如果最近的活动循环不是文本回复，或者没有活动循环
|
||||
cycle_info_block = "\n【近期回复历史】\n(最近没有连续文本回复)\n"
|
||||
|
||||
individuality = Individuality.get_instance()
|
||||
prompt_personality = individuality.get_prompt(x_person=2, level=2)
|
||||
|
||||
# 获取提示词模板并填充数据
|
||||
prompt = (await global_prompt_manager.get_prompt_async("planner_prompt")).format(
|
||||
bot_name=global_config.BOT_NICKNAME,
|
||||
prompt_personality=prompt_personality,
|
||||
structured_info_block=structured_info_block,
|
||||
chat_content_block=chat_content_block,
|
||||
current_mind_block=current_mind_block,
|
||||
@@ -1083,27 +1094,66 @@ class HeartFChatting:
|
||||
thinking_id: str,
|
||||
) -> Optional[List[str]]:
|
||||
"""
|
||||
回复器 (Replier): 核心逻辑用于生成回复。
|
||||
回复器 (Replier): 核心逻辑，负责生成回复文本。
|
||||
(已整合原 HeartFCGenerator 的功能)
|
||||
"""
|
||||
response_set: Optional[List[str]] = None
|
||||
try:
|
||||
response_set = await self.gpt_instance.generate_response(
|
||||
structured_info=self.sub_mind.structured_info,
|
||||
current_mind_info=self.sub_mind.current_mind,
|
||||
reason=reason,
|
||||
message=anchor_message, # Pass anchor_message positionally (matches 'message' parameter)
|
||||
thinking_id=thinking_id, # Pass thinking_id positionally
|
||||
)
|
||||
# 1. 获取情绪影响因子并调整模型温度
|
||||
arousal_multiplier = MoodManager.get_instance().get_arousal_multiplier()
|
||||
current_temp = global_config.llm_normal["temp"] * arousal_multiplier
|
||||
self.model_normal.temperature = current_temp # 动态调整温度
|
||||
|
||||
if not response_set:
|
||||
logger.warning(f"{self.log_prefix}[Replier-{thinking_id}] LLM生成了一个空回复集。")
|
||||
# 2. 获取信息捕捉器
|
||||
info_catcher = info_catcher_manager.get_info_catcher(thinking_id)
|
||||
|
||||
# 3. 构建 Prompt
|
||||
with Timer("构建Prompt", {}): # 内部计时器，可选保留
|
||||
prompt = await prompt_builder.build_prompt(
|
||||
build_mode="focus",
|
||||
reason=reason,
|
||||
current_mind_info=self.sub_mind.current_mind,
|
||||
structured_info=self.sub_mind.structured_info,
|
||||
message_txt="", # 似乎是固定的空字符串
|
||||
sender_name="", # 似乎是固定的空字符串
|
||||
chat_stream=anchor_message.chat_stream,
|
||||
)
|
||||
|
||||
# 4. 调用 LLM 生成回复
|
||||
content = None
|
||||
reasoning_content = None
|
||||
model_name = "unknown_model"
|
||||
try:
|
||||
with Timer("LLM生成", {}): # 内部计时器，可选保留
|
||||
content, reasoning_content, model_name = await self.model_normal.generate_response(prompt)
|
||||
logger.info(f"{self.log_prefix}[Replier-{thinking_id}]\\nPrompt:\\n{prompt}\\n生成回复: {content}\\n")
|
||||
# 捕捉 LLM 输出信息
|
||||
info_catcher.catch_after_llm_generated(
|
||||
prompt=prompt, response=content, reasoning_content=reasoning_content, model_name=model_name
|
||||
)
|
||||
|
||||
except Exception as llm_e:
|
||||
# 精简报错信息
|
||||
logger.error(f"{self.log_prefix}[Replier-{thinking_id}] LLM 生成失败: {llm_e}")
|
||||
return None # LLM 调用失败则无法生成回复
|
||||
|
||||
# 5. 处理 LLM 响应
|
||||
if not content:
|
||||
logger.warning(f"{self.log_prefix}[Replier-{thinking_id}] LLM 生成了空内容。")
|
||||
return None
|
||||
|
||||
return response_set
|
||||
with Timer("处理响应", {}): # 内部计时器，可选保留
|
||||
processed_response = process_llm_response(content)
|
||||
|
||||
if not processed_response:
|
||||
logger.warning(f"{self.log_prefix}[Replier-{thinking_id}] 处理后的回复为空。")
|
||||
return None
|
||||
|
||||
return processed_response
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"{self.log_prefix}[Replier-{thinking_id}] Unexpected error in replier_work: {e}")
|
||||
logger.error(traceback.format_exc())
|
||||
# 更通用的错误处理，精简信息
|
||||
logger.error(f"{self.log_prefix}[Replier-{thinking_id}] 回复生成意外失败: {e}")
|
||||
# logger.error(traceback.format_exc()) # 可以取消注释这行以在调试时查看完整堆栈
|
||||
return None
|
||||
|
||||
# --- Methods moved from HeartFCController start ---
|
||||
|
||||
@@ -1,107 +0,0 @@
|
||||
from typing import List, Optional
|
||||
|
||||
|
||||
from ..models.utils_model import LLMRequest
|
||||
from ...config.config import global_config
|
||||
from ..chat.message import MessageRecv
|
||||
from .heartflow_prompt_builder import prompt_builder
|
||||
from ..chat.utils import process_llm_response
|
||||
from src.common.logger_manager import get_logger
|
||||
from src.plugins.respon_info_catcher.info_catcher import info_catcher_manager
|
||||
from ..utils.timer_calculator import Timer
|
||||
|
||||
from src.plugins.moods.moods import MoodManager
|
||||
|
||||
|
||||
logger = get_logger("llm")
|
||||
|
||||
|
||||
class HeartFCGenerator:
|
||||
def __init__(self):
|
||||
self.model_normal = LLMRequest(
|
||||
model=global_config.llm_normal,
|
||||
temperature=global_config.llm_normal["temp"],
|
||||
max_tokens=256,
|
||||
request_type="response_heartflow",
|
||||
)
|
||||
|
||||
self.model_sum = LLMRequest(
|
||||
model=global_config.llm_summary_by_topic, temperature=0.6, max_tokens=2000, request_type="relation"
|
||||
)
|
||||
self.current_model_type = "r1" # 默认使用 R1
|
||||
self.current_model_name = "unknown model"
|
||||
|
||||
async def generate_response(
|
||||
self,
|
||||
structured_info: str,
|
||||
current_mind_info: str,
|
||||
reason: str,
|
||||
message: MessageRecv,
|
||||
thinking_id: str,
|
||||
) -> Optional[List[str]]:
|
||||
"""根据当前模型类型选择对应的生成函数"""
|
||||
|
||||
arousal_multiplier = MoodManager.get_instance().get_arousal_multiplier()
|
||||
|
||||
current_model = self.model_normal
|
||||
current_model.temperature = global_config.llm_normal["temp"] * arousal_multiplier # 激活度越高，温度越高
|
||||
model_response = await self._generate_response_with_model(
|
||||
structured_info, current_mind_info, reason, message, current_model, thinking_id
|
||||
)
|
||||
|
||||
if model_response:
|
||||
model_processed_response = await self._process_response(model_response)
|
||||
|
||||
return model_processed_response
|
||||
else:
|
||||
logger.info(f"{self.current_model_type}思考，失败")
|
||||
return None
|
||||
|
||||
async def _generate_response_with_model(
|
||||
self,
|
||||
structured_info: str,
|
||||
current_mind_info: str,
|
||||
reason: str,
|
||||
message: MessageRecv,
|
||||
model: LLMRequest,
|
||||
thinking_id: str,
|
||||
) -> str:
|
||||
info_catcher = info_catcher_manager.get_info_catcher(thinking_id)
|
||||
|
||||
with Timer() as _build_prompt:
|
||||
prompt = await prompt_builder.build_prompt(
|
||||
build_mode="focus",
|
||||
reason=reason,
|
||||
current_mind_info=current_mind_info,
|
||||
structured_info=structured_info,
|
||||
message_txt="",
|
||||
sender_name="",
|
||||
chat_stream=message.chat_stream,
|
||||
)
|
||||
# logger.info(f"构建prompt时间: {t_build_prompt.human_readable}")
|
||||
|
||||
try:
|
||||
content, reasoning_content, self.current_model_name = await model.generate_response(prompt)
|
||||
|
||||
logger.info(f"\nprompt:{prompt}\n生成回复{content}\n")
|
||||
|
||||
info_catcher.catch_after_llm_generated(
|
||||
prompt=prompt, response=content, reasoning_content=reasoning_content, model_name=self.current_model_name
|
||||
)
|
||||
|
||||
except Exception:
|
||||
logger.exception("生成回复时出错")
|
||||
return None
|
||||
|
||||
return content
|
||||
|
||||
async def _process_response(self, content: str) -> List[str]:
|
||||
"""处理响应内容，返回处理后的内容和情感标签"""
|
||||
if not content:
|
||||
return None
|
||||
|
||||
processed_response = process_llm_response(content)
|
||||
|
||||
# print(f"得到了处理后的llm返回{processed_response}")
|
||||
|
||||
return processed_response
|
||||
@@ -25,7 +25,7 @@ def init_prompt():
|
||||
{chat_target}
|
||||
{chat_talking_prompt}
|
||||
现在你想要在群里发言或者回复。\n
|
||||
你需要扮演一位网名叫{bot_name}的人进行回复，这个人的特点是："{prompt_personality} {prompt_identity}"。
|
||||
你需要扮演一位网名叫{bot_name}的人进行回复，这个人的特点是："{prompt_personality}"。
|
||||
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，你可以参考贴吧，知乎或者微博的回复风格。
|
||||
看到以上聊天记录，你刚刚在想：
|
||||
|
||||
@@ -49,7 +49,7 @@ def init_prompt():
|
||||
|
||||
# Planner提示词 - 优化版
|
||||
Prompt(
|
||||
"""你的名字是{bot_name}，一名大二女生，正在QQ群聊中。需要基于以下信息决定如何参与对话：
|
||||
"""你的名字是{bot_name},{prompt_personality}，你现在正在一个群聊中。需要基于以下信息决定如何参与对话：
|
||||
{structured_info_block}
|
||||
{chat_content_block}
|
||||
你的内心想法：
|
||||
@@ -164,9 +164,7 @@ class PromptBuilder:
|
||||
|
||||
async def _build_prompt_focus(self, reason, current_mind_info, structured_info, chat_stream) -> tuple[str, str]:
|
||||
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)
|
||||
|
||||
prompt_personality = individuality.get_prompt(x_person=0, level=2)
|
||||
# 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)}'''
|
||||
|
||||
@@ -214,7 +212,6 @@ class PromptBuilder:
|
||||
chat_talking_prompt=chat_talking_prompt,
|
||||
bot_name=global_config.BOT_NICKNAME,
|
||||
prompt_personality=prompt_personality,
|
||||
prompt_identity=prompt_identity,
|
||||
chat_target_2=await global_prompt_manager.get_prompt_async("chat_target_group2")
|
||||
if chat_in_group
|
||||
else await global_prompt_manager.get_prompt_async("chat_target_private2"),
|
||||
@@ -230,21 +227,8 @@ class PromptBuilder:
|
||||
return prompt
|
||||
|
||||
async def _build_prompt_normal(self, chat_stream, message_txt: str, sender_name: str = "某人") -> tuple[str, str]:
|
||||
# 开始构建prompt
|
||||
prompt_personality = "你"
|
||||
# person
|
||||
individuality = Individuality.get_instance()
|
||||
|
||||
personality_core = individuality.personality.personality_core
|
||||
prompt_personality += personality_core
|
||||
|
||||
personality_sides = individuality.personality.personality_sides
|
||||
random.shuffle(personality_sides)
|
||||
prompt_personality += f",{personality_sides[0]}"
|
||||
|
||||
identity_detail = individuality.identity.identity_detail
|
||||
random.shuffle(identity_detail)
|
||||
prompt_personality += f",{identity_detail[0]}"
|
||||
prompt_personality = individuality.get_prompt(x_person=2, level=2)
|
||||
|
||||
# 关系
|
||||
who_chat_in_group = [
|
||||
|
||||
@@ -14,51 +14,14 @@ from ...common.database import db
|
||||
from ...plugins.models.utils_model import LLMRequest
|
||||
from src.common.logger_manager import get_logger
|
||||
from src.plugins.memory_system.sample_distribution import MemoryBuildScheduler # 分布生成器
|
||||
from ..utils.chat_message_builder import (
|
||||
get_raw_msg_by_timestamp,
|
||||
build_readable_messages,
|
||||
) # 导入 build_readable_messages
|
||||
from ..chat.utils import translate_timestamp_to_human_readable
|
||||
from .memory_config import MemoryConfig
|
||||
|
||||
|
||||
def get_closest_chat_from_db(length: int, timestamp: str):
|
||||
# print(f"获取最接近指定时间戳的聊天记录，长度: {length}, 时间戳: {timestamp}")
|
||||
# print(f"当前时间: {timestamp},转换后时间: {time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(timestamp))}")
|
||||
chat_records = []
|
||||
closest_record = db.messages.find_one({"time": {"$lte": timestamp}}, sort=[("time", -1)])
|
||||
# print(f"最接近的记录: {closest_record}")
|
||||
if closest_record:
|
||||
closest_time = closest_record["time"]
|
||||
chat_id = closest_record["chat_id"] # 获取chat_id
|
||||
# 获取该时间戳之后的length条消息，保持相同的chat_id
|
||||
chat_records = list(
|
||||
db.messages.find(
|
||||
{
|
||||
"time": {"$gt": closest_time},
|
||||
"chat_id": chat_id, # 添加chat_id过滤
|
||||
}
|
||||
)
|
||||
.sort("time", 1)
|
||||
.limit(length)
|
||||
)
|
||||
# print(f"获取到的记录: {chat_records}")
|
||||
length = len(chat_records)
|
||||
# print(f"获取到的记录长度: {length}")
|
||||
# 转换记录格式
|
||||
formatted_records = []
|
||||
for record in chat_records:
|
||||
# 兼容行为，前向兼容老数据
|
||||
formatted_records.append(
|
||||
{
|
||||
"_id": record["_id"],
|
||||
"time": record["time"],
|
||||
"chat_id": record["chat_id"],
|
||||
"detailed_plain_text": record.get("detailed_plain_text", ""), # 添加文本内容
|
||||
"memorized_times": record.get("memorized_times", 0), # 添加记忆次数
|
||||
}
|
||||
)
|
||||
|
||||
return formatted_records
|
||||
|
||||
return []
|
||||
|
||||
|
||||
def calculate_information_content(text):
|
||||
"""计算文本的信息量（熵）"""
|
||||
char_count = Counter(text)
|
||||
@@ -263,16 +226,17 @@ class Hippocampus:
|
||||
@staticmethod
|
||||
def find_topic_llm(text, topic_num):
|
||||
prompt = (
|
||||
f"这是一段文字：{text}。请你从这段话中总结出最多{topic_num}个关键的概念，可以是名词，动词，或者特定人物，帮我列出来，"
|
||||
f"这是一段文字：\n{text}\n\n请你从这段话中总结出最多{topic_num}个关键的概念，可以是名词，动词，或者特定人物，帮我列出来，"
|
||||
f"将主题用逗号隔开，并加上<>,例如<主题1>,<主题2>......尽可能精简。只需要列举最多{topic_num}个话题就好，不要有序号，不要告诉我其他内容。"
|
||||
f"如果确定找不出主题或者没有明显主题，返回<none>。"
|
||||
)
|
||||
return prompt
|
||||
|
||||
@staticmethod
|
||||
def topic_what(text, topic, time_info):
|
||||
def topic_what(text, topic):
|
||||
# 不再需要 time_info 参数
|
||||
prompt = (
|
||||
f'这是一段文字，{time_info}：{text}。我想让你基于这段文字来概括"{topic}"这个概念，帮我总结成一句自然的话，'
|
||||
f'这是一段文字：\n{text}\n\n我想让你基于这段文字来概括"{topic}"这个概念，帮我总结成一句自然的话，'
|
||||
f"可以包含时间和人物，以及具体的观点。只输出这句话就好"
|
||||
)
|
||||
return prompt
|
||||
@@ -845,9 +809,12 @@ class EntorhinalCortex:
|
||||
)
|
||||
|
||||
timestamps = sample_scheduler.get_timestamp_array()
|
||||
logger.info(f"回忆往事: {[time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(ts)) for ts in timestamps]}")
|
||||
# 使用 translate_timestamp_to_human_readable 并指定 mode="normal"
|
||||
readable_timestamps = [translate_timestamp_to_human_readable(ts, mode="normal") for ts in timestamps]
|
||||
logger.info(f"回忆往事: {readable_timestamps}")
|
||||
chat_samples = []
|
||||
for timestamp in timestamps:
|
||||
# 调用修改后的 random_get_msg_snippet
|
||||
messages = self.random_get_msg_snippet(
|
||||
timestamp, self.config.build_memory_sample_length, max_memorized_time_per_msg
|
||||
)
|
||||
@@ -862,22 +829,45 @@ class EntorhinalCortex:
|
||||
|
||||
@staticmethod
|
||||
def random_get_msg_snippet(target_timestamp: float, chat_size: int, max_memorized_time_per_msg: int) -> list:
|
||||
"""从数据库中随机获取指定时间戳附近的消息片段"""
|
||||
"""从数据库中随机获取指定时间戳附近的消息片段 (使用 chat_message_builder)"""
|
||||
try_count = 0
|
||||
time_window_seconds = random.randint(300, 1800) # 随机时间窗口，5到30分钟
|
||||
|
||||
while try_count < 3:
|
||||
messages = get_closest_chat_from_db(length=chat_size, timestamp=target_timestamp)
|
||||
# 定义时间范围：从目标时间戳开始，向后推移 time_window_seconds
|
||||
timestamp_start = target_timestamp
|
||||
timestamp_end = target_timestamp + time_window_seconds
|
||||
|
||||
# 使用 chat_message_builder 的函数获取消息
|
||||
# limit_mode='earliest' 获取这个时间窗口内最早的 chat_size 条消息
|
||||
messages = get_raw_msg_by_timestamp(
|
||||
timestamp_start=timestamp_start, timestamp_end=timestamp_end, limit=chat_size, limit_mode="earliest"
|
||||
)
|
||||
|
||||
if messages:
|
||||
# 检查获取到的所有消息是否都未达到最大记忆次数
|
||||
all_valid = True
|
||||
for message in messages:
|
||||
if message["memorized_times"] >= max_memorized_time_per_msg:
|
||||
messages = None
|
||||
if message.get("memorized_times", 0) >= max_memorized_time_per_msg:
|
||||
all_valid = False
|
||||
break
|
||||
if messages:
|
||||
|
||||
# 如果所有消息都有效
|
||||
if all_valid:
|
||||
# 更新数据库中的记忆次数
|
||||
for message in messages:
|
||||
# 确保在更新前获取最新的 memorized_times，以防万一
|
||||
current_memorized_times = message.get("memorized_times", 0)
|
||||
db.messages.update_one(
|
||||
{"_id": message["_id"]}, {"$set": {"memorized_times": message["memorized_times"] + 1}}
|
||||
{"_id": message["_id"]}, {"$set": {"memorized_times": current_memorized_times + 1}}
|
||||
)
|
||||
return messages
|
||||
return messages # 直接返回原始的消息列表
|
||||
|
||||
# 如果获取失败或消息无效，增加尝试次数
|
||||
try_count += 1
|
||||
target_timestamp -= 120 # 如果第一次尝试失败，稍微向前调整时间戳再试
|
||||
|
||||
# 三次尝试都失败，返回 None
|
||||
return None
|
||||
|
||||
async def sync_memory_to_db(self):
|
||||
@@ -1113,86 +1103,70 @@ class ParahippocampalGyrus:
|
||||
"""压缩和总结消息内容，生成记忆主题和摘要。
|
||||
|
||||
Args:
|
||||
messages (list): 消息列表，每个消息是一个字典，包含以下字段：
|
||||
- time: float, 消息的时间戳
|
||||
- detailed_plain_text: str, 消息的详细文本内容
|
||||
messages (list): 消息列表，每个消息是一个字典，包含数据库消息结构。
|
||||
compress_rate (float, optional): 压缩率，用于控制生成的主题数量。默认为0.1。
|
||||
|
||||
Returns:
|
||||
tuple: (compressed_memory, similar_topics_dict)
|
||||
- compressed_memory: set, 压缩后的记忆集合，每个元素是一个元组 (topic, summary)
|
||||
- topic: str, 记忆主题
|
||||
- summary: str, 主题的摘要描述
|
||||
- similar_topics_dict: dict, 相似主题字典，key为主题，value为相似主题列表
|
||||
每个相似主题是一个元组 (similar_topic, similarity)
|
||||
- similar_topic: str, 相似的主题
|
||||
- similarity: float, 相似度分数（0-1之间）
|
||||
- similar_topics_dict: dict, 相似主题字典
|
||||
|
||||
Process:
|
||||
1. 合并消息文本并生成时间信息
|
||||
2. 使用LLM提取关键主题
|
||||
3. 过滤掉包含禁用关键词的主题
|
||||
4. 为每个主题生成摘要
|
||||
5. 查找与现有记忆中的相似主题
|
||||
1. 使用 build_readable_messages 生成包含时间、人物信息的格式化文本。
|
||||
2. 使用LLM提取关键主题。
|
||||
3. 过滤掉包含禁用关键词的主题。
|
||||
4. 为每个主题生成摘要。
|
||||
5. 查找与现有记忆中的相似主题。
|
||||
"""
|
||||
if not messages:
|
||||
return set(), {}
|
||||
|
||||
# 合并消息文本，同时保留时间信息
|
||||
input_text = ""
|
||||
time_info = ""
|
||||
# 计算最早和最晚时间
|
||||
earliest_time = min(msg["time"] for msg in messages)
|
||||
latest_time = max(msg["time"] for msg in messages)
|
||||
# 1. 使用 build_readable_messages 生成格式化文本
|
||||
# build_readable_messages 只返回一个字符串，不需要解包
|
||||
input_text = await build_readable_messages(
|
||||
messages,
|
||||
merge_messages=True, # 合并连续消息
|
||||
timestamp_mode="normal", # 使用 'YYYY-MM-DD HH:MM:SS' 格式
|
||||
replace_bot_name=False, # 保留原始用户名
|
||||
)
|
||||
|
||||
earliest_dt = datetime.datetime.fromtimestamp(earliest_time)
|
||||
latest_dt = datetime.datetime.fromtimestamp(latest_time)
|
||||
# 如果生成的可读文本为空（例如所有消息都无效），则直接返回
|
||||
if not input_text:
|
||||
logger.warning("无法从提供的消息生成可读文本，跳过记忆压缩。")
|
||||
return set(), {}
|
||||
|
||||
# 如果是同一年
|
||||
if earliest_dt.year == latest_dt.year:
|
||||
earliest_str = earliest_dt.strftime("%m-%d %H:%M:%S")
|
||||
latest_str = latest_dt.strftime("%m-%d %H:%M:%S")
|
||||
time_info += f"是在{earliest_dt.year}年，{earliest_str} 到 {latest_str} 的对话:\n"
|
||||
else:
|
||||
earliest_str = earliest_dt.strftime("%Y-%m-%d %H:%M:%S")
|
||||
latest_str = latest_dt.strftime("%Y-%m-%d %H:%M:%S")
|
||||
time_info += f"是从 {earliest_str} 到 {latest_str} 的对话:\n"
|
||||
|
||||
for msg in messages:
|
||||
input_text += f"{msg['detailed_plain_text']}\n"
|
||||
|
||||
logger.debug(input_text)
|
||||
logger.debug(f"用于压缩的格式化文本:\n{input_text}")
|
||||
|
||||
# 2. 使用LLM提取关键主题
|
||||
topic_num = self.hippocampus.calculate_topic_num(input_text, compress_rate)
|
||||
topics_response = await self.hippocampus.llm_topic_judge.generate_response(
|
||||
self.hippocampus.find_topic_llm(input_text, topic_num)
|
||||
)
|
||||
|
||||
# 使用正则表达式提取<>中的内容
|
||||
# 提取<>中的内容
|
||||
topics = re.findall(r"<([^>]+)>", topics_response[0])
|
||||
|
||||
# 如果没有找到<>包裹的内容，返回['none']
|
||||
if not topics:
|
||||
topics = ["none"]
|
||||
else:
|
||||
# 处理提取出的话题
|
||||
topics = [
|
||||
topic.strip()
|
||||
for topic in ",".join(topics).replace("，", ",").replace("、", ",").replace(" ", ",").split(",")
|
||||
if topic.strip()
|
||||
]
|
||||
|
||||
# 过滤掉包含禁用关键词的topic
|
||||
# 3. 过滤掉包含禁用关键词的topic
|
||||
filtered_topics = [
|
||||
topic for topic in topics if not any(keyword in topic for keyword in self.config.memory_ban_words)
|
||||
]
|
||||
|
||||
logger.debug(f"过滤后话题: {filtered_topics}")
|
||||
|
||||
# 创建所有话题的请求任务
|
||||
# 4. 创建所有话题的摘要生成任务
|
||||
tasks = []
|
||||
for topic in filtered_topics:
|
||||
topic_what_prompt = self.hippocampus.topic_what(input_text, topic, time_info)
|
||||
# 调用修改后的 topic_what，不再需要 time_info
|
||||
topic_what_prompt = self.hippocampus.topic_what(input_text, topic)
|
||||
try:
|
||||
task = self.hippocampus.llm_summary_by_topic.generate_response_async(topic_what_prompt)
|
||||
tasks.append((topic.strip(), task))
|
||||
|
||||
@@ -750,7 +750,6 @@ class LLMRequest:
|
||||
"tools": tools,
|
||||
}
|
||||
|
||||
|
||||
response = await self._execute_request(endpoint="/chat/completions", payload=data, prompt=prompt)
|
||||
logger.debug(f"向模型 {self.model_name} 发送工具调用请求，包含 {len(tools)} 个工具，返回结果: {response}")
|
||||
# 检查响应是否包含工具调用
|
||||
|
||||
@@ -180,10 +180,10 @@ class PersonInfoManager:
|
||||
existing_names = ""
|
||||
while current_try < max_retries:
|
||||
individuality = Individuality.get_instance()
|
||||
prompt_personality = individuality.get_prompt(type="personality", x_person=2, level=1)
|
||||
prompt_personality = individuality.get_prompt(x_person=2, level=1)
|
||||
bot_name = individuality.personality.bot_nickname
|
||||
|
||||
qv_name_prompt = f"你是{bot_name}，你{prompt_personality}"
|
||||
qv_name_prompt = f"你是{bot_name}，{prompt_personality}"
|
||||
qv_name_prompt += f"现在你想给一个用户取一个昵称，用户是的qq昵称是{user_nickname}，"
|
||||
qv_name_prompt += f"用户的qq群昵称名是{user_cardname}，"
|
||||
if user_avatar:
|
||||
|
||||
@@ -1,6 +1,6 @@
|
||||
import json
|
||||
import logging
|
||||
from typing import Any, Dict, TypeVar, List, Union, Callable, Tuple
|
||||
from typing import Any, Dict, TypeVar, List, Union, Tuple
|
||||
|
||||
# 定义类型变量用于泛型类型提示
|
||||
T = TypeVar("T")
|
||||
@@ -70,7 +70,6 @@ def extract_tool_call_arguments(tool_call: Dict[str, Any], default_value: Dict[s
|
||||
return default_result
|
||||
|
||||
|
||||
|
||||
def safe_json_dumps(obj: Any, default_value: str = "{}", ensure_ascii: bool = False, pretty: bool = False) -> str:
|
||||
"""
|
||||
安全地将Python对象序列化为JSON字符串
|
||||
@@ -95,8 +94,6 @@ def safe_json_dumps(obj: Any, default_value: str = "{}", ensure_ascii: bool = Fa
|
||||
return default_value
|
||||
|
||||
|
||||
|
||||
|
||||
def normalize_llm_response(response: Any, log_prefix: str = "") -> Tuple[bool, List[Any], str]:
|
||||
"""
|
||||
标准化LLM响应格式，将各种格式（如元组）转换为统一的列表格式
|
||||
@@ -108,9 +105,9 @@ def normalize_llm_response(response: Any, log_prefix: str = "") -> Tuple[bool, L
|
||||
返回:
|
||||
元组 (成功标志, 标准化后的响应列表, 错误消息)
|
||||
"""
|
||||
|
||||
|
||||
logger.debug(f"{log_prefix}原始人 LLM响应: {response}")
|
||||
|
||||
|
||||
# 检查是否为None
|
||||
if response is None:
|
||||
return False, [], "LLM响应为None"
|
||||
@@ -140,7 +137,9 @@ def normalize_llm_response(response: Any, log_prefix: str = "") -> Tuple[bool, L
|
||||
return True, response, ""
|
||||
|
||||
|
||||
def process_llm_tool_calls(tool_calls: List[Dict[str, Any]], log_prefix: str = "") -> Tuple[bool, List[Dict[str, Any]], str]:
|
||||
def process_llm_tool_calls(
|
||||
tool_calls: List[Dict[str, Any]], log_prefix: str = ""
|
||||
) -> Tuple[bool, List[Dict[str, Any]], str]:
|
||||
"""
|
||||
处理并验证LLM响应中的工具调用列表
|
||||
|
||||
@@ -165,7 +164,9 @@ def process_llm_tool_calls(tool_calls: List[Dict[str, Any]], log_prefix: str = "
|
||||
|
||||
# 检查基本结构
|
||||
if tool_call.get("type") != "function":
|
||||
logger.warning(f"{log_prefix}工具调用[{i}]不是function类型: type={tool_call.get('type', '未定义')}, 内容: {tool_call}")
|
||||
logger.warning(
|
||||
f"{log_prefix}工具调用[{i}]不是function类型: type={tool_call.get('type', '未定义')}, 内容: {tool_call}"
|
||||
)
|
||||
continue
|
||||
|
||||
if "function" not in tool_call or not isinstance(tool_call.get("function"), dict):
|
||||
@@ -176,16 +177,20 @@ def process_llm_tool_calls(tool_calls: List[Dict[str, Any]], log_prefix: str = "
|
||||
if "name" not in func_details or not isinstance(func_details.get("name"), str):
|
||||
logger.warning(f"{log_prefix}工具调用[{i}]的'function'字段缺少'name'或类型不正确: {func_details}")
|
||||
continue
|
||||
if "arguments" not in func_details or not isinstance(func_details.get("arguments"), str): # 参数是字符串形式的JSON
|
||||
if "arguments" not in func_details or not isinstance(
|
||||
func_details.get("arguments"), str
|
||||
): # 参数是字符串形式的JSON
|
||||
logger.warning(f"{log_prefix}工具调用[{i}]的'function'字段缺少'arguments'或类型不正确: {func_details}")
|
||||
continue
|
||||
|
||||
# 可选：尝试解析参数JSON，确保其有效
|
||||
args_str = func_details["arguments"]
|
||||
try:
|
||||
json.loads(args_str) # 尝试解析，但不存储结果
|
||||
json.loads(args_str) # 尝试解析，但不存储结果
|
||||
except json.JSONDecodeError as e:
|
||||
logger.warning(f"{log_prefix}工具调用[{i}]的'arguments'不是有效的JSON字符串: {e}, 内容: {args_str[:100]}...")
|
||||
logger.warning(
|
||||
f"{log_prefix}工具调用[{i}]的'arguments'不是有效的JSON字符串: {e}, 内容: {args_str[:100]}..."
|
||||
)
|
||||
continue
|
||||
except Exception as e:
|
||||
logger.warning(f"{log_prefix}解析工具调用[{i}]的'arguments'时发生意外错误: {e}, 内容: {args_str[:100]}...")
|
||||
@@ -193,7 +198,7 @@ def process_llm_tool_calls(tool_calls: List[Dict[str, Any]], log_prefix: str = "
|
||||
|
||||
valid_tool_calls.append(tool_call)
|
||||
|
||||
if not valid_tool_calls and tool_calls: # 如果原始列表不为空，但验证后为空
|
||||
if not valid_tool_calls and tool_calls: # 如果原始列表不为空，但验证后为空
|
||||
return False, [], "所有工具调用格式均无效"
|
||||
|
||||
return True, valid_tool_calls, ""
|
||||
|
||||
Reference in New Issue
Block a user