mirror of https://github.com/Mai-with-u/MaiBot.git
313 lines
16 KiB
Python
313 lines
16 KiB
Python
from typing import List, Dict, Any
|
||
from .chat_observer import ChatObserver
|
||
from src.common.logger_manager import get_logger
|
||
from src.chat.models.utils_model import LLMRequest
|
||
from src.chat.person_info.person_info import person_info_manager
|
||
from src.chat.person_info.relationship_manager import relationship_manager
|
||
from src.chat.utils.chat_message_builder import build_readable_messages
|
||
from .observation_info import ObservationInfo
|
||
from .conversation_info import ConversationInfo
|
||
from .pfc_utils import get_items_from_json, adjust_relationship_value_nonlinear
|
||
from src.config.config import global_config # 导入全局配置 (向上两级到 src/, 再到 config)
|
||
|
||
|
||
logger = get_logger("pfc_relationship")
|
||
|
||
|
||
class PfcRelationshipUpdater:
|
||
def __init__(self, private_name: str, bot_name: str):
|
||
"""
|
||
初始化关系更新器。
|
||
|
||
Args:
|
||
private_name (str): 当前私聊对象的名称 (用于日志)。
|
||
bot_name (str): 机器人自己的名称。
|
||
"""
|
||
self.private_name = private_name
|
||
self.bot_name = bot_name
|
||
self.person_info_mng = person_info_manager
|
||
self.relationship_mng = relationship_manager # 复用其实例方法
|
||
|
||
# LLM 实例 (为关系评估创建一个新的)
|
||
# 尝试读取 llm_PFC_relationship_eval 配置，如果不存在则回退
|
||
llm_config_rel_eval = getattr(global_config, "llm_PFC_relationship_eval", None)
|
||
if llm_config_rel_eval and isinstance(llm_config_rel_eval, dict):
|
||
logger.info(f"[私聊][{self.private_name}] 使用 llm_PFC_relationship_eval 配置初始化关系评估LLM。")
|
||
self.llm = LLMRequest(
|
||
model=llm_config_rel_eval,
|
||
temperature=llm_config_rel_eval.get("temp", 0.5), # 判断任务通常用较低温度
|
||
max_tokens=llm_config_rel_eval.get("max_tokens", 512),
|
||
request_type="pfc_relationship_evaluation",
|
||
)
|
||
else:
|
||
logger.warning(
|
||
f"[私聊][{self.private_name}] 未找到 llm_PFC_relationship_eval 配置或配置无效，将回退使用 llm_PFC_action_planner 的配置。"
|
||
)
|
||
llm_config_action_planner = getattr(global_config, "llm_PFC_action_planner", None)
|
||
if llm_config_action_planner and isinstance(llm_config_action_planner, dict):
|
||
self.llm = LLMRequest(
|
||
model=llm_config_action_planner, # 使用 action_planner 的模型配置
|
||
temperature=llm_config_action_planner.get("temp", 0.5), # 但温度可以尝试低一些
|
||
max_tokens=llm_config_action_planner.get("max_tokens", 512),
|
||
request_type="pfc_relationship_evaluation_fallback",
|
||
)
|
||
else: # 极端情况，连 action_planner 的配置都没有
|
||
logger.error(f"[私聊][{self.private_name}] 无法找到任何有效的LLM配置用于关系评估！关系更新功能将受限。")
|
||
self.llm = None # LLM 未初始化
|
||
|
||
# 从 global_config 读取参数，若无则使用默认值
|
||
self.REL_INCREMENTAL_INTERVAL = getattr(global_config, "pfc_relationship_incremental_interval", 10)
|
||
self.REL_INCREMENTAL_MSG_COUNT = getattr(global_config, "pfc_relationship_incremental_msg_count", 10)
|
||
self.REL_INCREMENTAL_DEFAULT_CHANGE = getattr(global_config, "pfc_relationship_incremental_default_change", 1.0)
|
||
self.REL_INCREMENTAL_MAX_CHANGE = getattr(global_config, "pfc_relationship_incremental_max_change", 5.0)
|
||
|
||
self.REL_FINAL_MSG_COUNT = getattr(global_config, "pfc_relationship_final_msg_count", 30)
|
||
self.REL_FINAL_DEFAULT_CHANGE = getattr(global_config, "pfc_relationship_final_default_change", 5.0)
|
||
self.REL_FINAL_MAX_CHANGE = getattr(global_config, "pfc_relationship_final_max_change", 50.0)
|
||
|
||
async def update_relationship_incremental(
|
||
self,
|
||
conversation_info: ConversationInfo,
|
||
observation_info: ObservationInfo,
|
||
chat_observer_for_history: ChatObserver, # ChatObserver 实例
|
||
) -> None:
|
||
if not self.llm:
|
||
logger.error(f"[私聊][{self.private_name}] LLM未初始化，无法进行增量关系更新。")
|
||
return
|
||
if not conversation_info or not conversation_info.person_id or not observation_info:
|
||
logger.debug(f"[私聊][{self.private_name}] 增量关系更新：缺少必要信息。")
|
||
return
|
||
|
||
if not (
|
||
conversation_info.current_instance_message_count % self.REL_INCREMENTAL_INTERVAL == 0
|
||
and conversation_info.current_instance_message_count > 0
|
||
):
|
||
return
|
||
|
||
logger.info(
|
||
f"[私聊][{self.private_name}] 达到增量关系更新阈值 ({conversation_info.current_instance_message_count}条消息)，开始评估..."
|
||
)
|
||
|
||
messages_for_eval: List[Dict[str, Any]] = []
|
||
if chat_observer_for_history:
|
||
messages_for_eval = chat_observer_for_history.get_cached_messages(limit=self.REL_INCREMENTAL_MSG_COUNT)
|
||
elif observation_info.chat_history:
|
||
messages_for_eval = observation_info.chat_history[-self.REL_INCREMENTAL_MSG_COUNT :]
|
||
|
||
if not messages_for_eval:
|
||
logger.warning(f"[私聊][{self.private_name}] 增量关系更新：没有足够的消息进行评估。")
|
||
return
|
||
|
||
readable_history_for_llm = await build_readable_messages(
|
||
messages_for_eval, replace_bot_name=True, merge_messages=False, timestamp_mode="relative"
|
||
)
|
||
|
||
current_relationship_value = await self.person_info_mng.get_value(
|
||
conversation_info.person_id, "relationship_value"
|
||
)
|
||
current_relationship_value = self.relationship_mng.ensure_float(
|
||
current_relationship_value, conversation_info.person_id
|
||
)
|
||
|
||
relationship_prompt = f"""你是{self.bot_name}。你正在与{self.private_name}私聊。
|
||
你们当前的关系值大约是 {current_relationship_value:.0f} (范围通常在-1000到1000，越高越代表关系越好)。
|
||
以下是你们最近的对话内容：
|
||
---
|
||
{readable_history_for_llm}
|
||
---
|
||
请基于以上对话，判断你与{self.private_name}的关系值应该如何“谨慎地”调整。
|
||
请输出一个JSON对象，包含一个 "adjustment" 字段，其值为一个介于 -{self.REL_INCREMENTAL_MAX_CHANGE} 和 +{self.REL_INCREMENTAL_MAX_CHANGE} 之间的整数，代表关系值的变化。
|
||
例如：{{ "adjustment": 3 }}。如果对话内容不明确或难以判断，请倾向于输出较小的调整值（如0, 1, -1）。"""
|
||
|
||
raw_adjustment_val = self.REL_INCREMENTAL_DEFAULT_CHANGE
|
||
try:
|
||
logger.debug(f"[私聊][{self.private_name}] 增量关系评估Prompt:\n{relationship_prompt}")
|
||
content, _ = await self.llm.generate_response_async(relationship_prompt)
|
||
logger.debug(f"[私聊][{self.private_name}] 增量关系评估LLM原始返回: {content}")
|
||
|
||
success, result = get_items_from_json(
|
||
content,
|
||
self.private_name,
|
||
"adjustment",
|
||
default_values={"adjustment": self.REL_INCREMENTAL_DEFAULT_CHANGE},
|
||
required_types={"adjustment": (int, float)},
|
||
)
|
||
raw_adjustment = result.get("adjustment", self.REL_INCREMENTAL_DEFAULT_CHANGE)
|
||
if not isinstance(raw_adjustment, (int, float)):
|
||
raw_adjustment_val = self.REL_INCREMENTAL_DEFAULT_CHANGE
|
||
else:
|
||
raw_adjustment_val = float(raw_adjustment)
|
||
raw_adjustment_val = max(
|
||
-self.REL_INCREMENTAL_MAX_CHANGE, min(self.REL_INCREMENTAL_MAX_CHANGE, raw_adjustment_val)
|
||
)
|
||
except Exception as e:
|
||
logger.error(f"[私聊][{self.private_name}] 增量关系评估LLM调用或解析失败: {e}")
|
||
|
||
adjustment_val = await adjust_relationship_value_nonlinear(current_relationship_value, raw_adjustment_val)
|
||
|
||
new_relationship_value = max(-1000.0, min(1000.0, current_relationship_value + adjustment_val))
|
||
await self.person_info_mng.update_one_field(
|
||
conversation_info.person_id, "relationship_value", new_relationship_value
|
||
)
|
||
logger.info(
|
||
f"[私聊][{self.private_name}] 增量关系值更新：与【{self.private_name}】的关系值从 {current_relationship_value:.2f} 调整了 {adjustment_val:.2f}，变为 {new_relationship_value:.2f}"
|
||
)
|
||
|
||
if conversation_info.person_id:
|
||
conversation_info.relationship_text = await self.relationship_mng.build_relationship_info(
|
||
conversation_info.person_id, is_id=True
|
||
)
|
||
|
||
async def update_relationship_final(
|
||
self,
|
||
conversation_info: ConversationInfo,
|
||
observation_info: ObservationInfo,
|
||
chat_observer_for_history: ChatObserver,
|
||
) -> None:
|
||
if not self.llm:
|
||
logger.error(f"[私聊][{self.private_name}] LLM未初始化，无法进行最终关系更新。")
|
||
return
|
||
if not conversation_info or not conversation_info.person_id or not observation_info:
|
||
logger.debug(f"[私聊][{self.private_name}] 最终关系更新：缺少必要信息。")
|
||
return
|
||
|
||
logger.info(f"[私聊][{self.private_name}] 私聊结束，开始最终关系评估...")
|
||
|
||
messages_for_eval: List[Dict[str, Any]] = []
|
||
if chat_observer_for_history:
|
||
messages_for_eval = chat_observer_for_history.get_cached_messages(limit=self.REL_FINAL_MSG_COUNT)
|
||
elif observation_info.chat_history:
|
||
messages_for_eval = observation_info.chat_history[-self.REL_FINAL_MSG_COUNT :]
|
||
|
||
if not messages_for_eval:
|
||
logger.warning(f"[私聊][{self.private_name}] 最终关系更新：没有足够的消息进行评估。")
|
||
return
|
||
|
||
readable_history_for_llm = await build_readable_messages(
|
||
messages_for_eval, replace_bot_name=True, merge_messages=False, timestamp_mode="relative"
|
||
)
|
||
|
||
current_relationship_value = await self.person_info_mng.get_value(
|
||
conversation_info.person_id, "relationship_value"
|
||
)
|
||
current_relationship_value = self.relationship_mng.ensure_float(
|
||
current_relationship_value, conversation_info.person_id
|
||
)
|
||
|
||
relationship_prompt = f"""你是{self.bot_name}。你与{self.private_name}的私聊刚刚结束。
|
||
你们当前的关系值大约是 {current_relationship_value:.0f} (范围通常在-1000到1000，越高越好)。
|
||
以下是你们本次私聊最后部分的对话内容：
|
||
---
|
||
{readable_history_for_llm}
|
||
---
|
||
请基于以上对话的整体情况，判断你与【{self.private_name}】的关系值应该如何进行一次总结性的调整。
|
||
请输出一个JSON对象，包含一个 "final_adjustment" 字段，其值为一个整数，代表关系值的变化量（例如，可以是 -{self.REL_FINAL_MAX_CHANGE} 到 +{self.REL_FINAL_MAX_CHANGE} 之间的一个值）。
|
||
请大胆评估，但也要合理。"""
|
||
|
||
raw_adjustment_val = self.REL_FINAL_DEFAULT_CHANGE
|
||
try:
|
||
logger.debug(f"[私聊][{self.private_name}] 最终关系评估Prompt:\n{relationship_prompt}")
|
||
content, _ = await self.llm.generate_response_async(relationship_prompt)
|
||
logger.debug(f"[私聊][{self.private_name}] 最终关系评估LLM原始返回: {content}")
|
||
|
||
success, result = get_items_from_json(
|
||
content,
|
||
self.private_name,
|
||
"final_adjustment",
|
||
default_values={"final_adjustment": self.REL_FINAL_DEFAULT_CHANGE},
|
||
required_types={"final_adjustment": (int, float)},
|
||
)
|
||
raw_adjustment = result.get("final_adjustment", self.REL_FINAL_DEFAULT_CHANGE)
|
||
if not isinstance(raw_adjustment, (int, float)):
|
||
raw_adjustment_val = self.REL_FINAL_DEFAULT_CHANGE
|
||
else:
|
||
raw_adjustment_val = float(raw_adjustment)
|
||
raw_adjustment_val = max(
|
||
-self.REL_INCREMENTAL_MAX_CHANGE, min(self.REL_INCREMENTAL_MAX_CHANGE, raw_adjustment_val)
|
||
)
|
||
except Exception as e:
|
||
logger.error(f"[私聊][{self.private_name}] 最终关系评估LLM调用或解析失败: {e}")
|
||
|
||
adjustment_val = await adjust_relationship_value_nonlinear(current_relationship_value, raw_adjustment_val)
|
||
|
||
new_relationship_value = max(-1000.0, min(1000.0, current_relationship_value + adjustment_val))
|
||
await self.person_info_mng.update_one_field(
|
||
conversation_info.person_id, "relationship_value", new_relationship_value
|
||
)
|
||
logger.info(
|
||
f"[私聊][{self.private_name}] 最终关系值更新：与【{self.private_name}】的关系值从 {current_relationship_value:.2f} 调整了 {adjustment_val:.2f}，最终为 {new_relationship_value:.2f}"
|
||
)
|
||
|
||
if conversation_info.person_id: # 虽然通常结束了，但更新一下无妨
|
||
conversation_info.relationship_text = await self.relationship_mng.build_relationship_info(
|
||
conversation_info.person_id, is_id=True
|
||
)
|
||
|
||
|
||
class PfcRepationshipTranslator:
|
||
"""直接完整导入群聊的relationship_manager.py可能不可取
|
||
因为对于PFC的planner来说
|
||
其暗示了选择回复
|
||
所以新建代码文件来适配PFC的决策层面"""
|
||
|
||
def __init__(self, private_name: str):
|
||
self.private_name = private_name
|
||
|
||
async def translate_relationship_value_to_text(self, relationship_value: float) -> str:
|
||
"""
|
||
将数值型的关系值转换为PFC私聊场景下简洁的关系描述文本。
|
||
"""
|
||
level_num = self._calculate_relationship_level_num(relationship_value, self.private_name)
|
||
|
||
relationship_descriptions = [
|
||
"厌恶", # level_num 0
|
||
"冷漠", # level_num 1
|
||
"初识", # level_num 2
|
||
"友好", # level_num 3
|
||
"喜欢", # level_num 4
|
||
"依赖", # level_num 5
|
||
]
|
||
|
||
if 0 <= level_num < len(relationship_descriptions):
|
||
description = relationship_descriptions[level_num]
|
||
else:
|
||
description = "普通" # 默认或错误情况
|
||
logger.warning(f"[私聊][{self.private_name}] 计算出的 level_num ({level_num}) 无效，关系描述默认为 '普通'")
|
||
|
||
return f"{description}。"
|
||
|
||
@staticmethod
|
||
def _calculate_relationship_level_num(relationship_value: float, private_name: str) -> int:
|
||
"""
|
||
根据关系值计算关系等级编号 (0-5)。
|
||
这里的阈值应与 relationship_manager.py 中的保持一致
|
||
"""
|
||
if not isinstance(relationship_value, (int, float)):
|
||
logger.warning(
|
||
f"[私聊][{private_name}] 传入的 relationship_value '{relationship_value}' 不是有效的数值类型，默认为0。"
|
||
)
|
||
relationship_value = 0.0
|
||
|
||
if -1000 <= relationship_value < -227:
|
||
level_num = 0 # 厌恶
|
||
elif -227 <= relationship_value < -73:
|
||
level_num = 1 # 冷漠
|
||
elif -73 <= relationship_value < 227:
|
||
level_num = 2 # 普通/认识
|
||
elif 227 <= relationship_value < 587:
|
||
level_num = 3 # 友好
|
||
elif 587 <= relationship_value < 900:
|
||
level_num = 4 # 喜欢
|
||
elif 900 <= relationship_value <= 1000:
|
||
level_num = 5 # 暧昧
|
||
else:
|
||
# 超出范围的值处理
|
||
if relationship_value > 1000:
|
||
level_num = 5
|
||
elif relationship_value < -1000:
|
||
level_num = 0
|
||
else: # 理论上不会到这里，除非前面的条件逻辑有误
|
||
logger.warning(f"[私聊][{private_name}] 关系值 {relationship_value} 未落入任何预设范围，默认为普通。")
|
||
level_num = 2
|
||
return level_num
|