mirror of https://github.com/Mai-with-u/MaiBot.git
人设详细程度，日程模型选择，心流去重可配置
Bakadax
parent
429eaaae4f
commit
68b3b11023
|
|
@ -153,6 +153,7 @@ class BotConfig:
|
|||
"用一句话或几句话描述人格的一些侧面",
|
||||
]
|
||||
)
|
||||
personality_detail_level: int = 0 # 人设消息注入 prompt 详细等级 (0: 采用默认配置, 1: 核心/随机细节, 2: 核心+随机侧面/全部细节, 3: 全部)
|
||||
# identity
|
||||
identity_detail: List[str] = field(
|
||||
default_factory=lambda: [
|
||||
|
|
@ -178,6 +179,7 @@ class BotConfig:
|
|||
|
||||
base_normal_chat_num: int = 3 # 最多允许多少个群进行普通聊天
|
||||
base_focused_chat_num: int = 2 # 最多允许多少个群进行专注聊天
|
||||
allow_remove_duplicates: bool = True # 是否开启心流去重（如果发现心流截断问题严重可尝试关闭）
|
||||
|
||||
observation_context_size: int = 12 # 心流观察到的最长上下文大小，超过这个值的上下文会被压缩
|
||||
|
||||
|
|
@ -296,7 +298,9 @@ class BotConfig:
|
|||
llm_heartflow: Dict[str, str] = field(default_factory=lambda: {})
|
||||
llm_tool_use: Dict[str, str] = field(default_factory=lambda: {})
|
||||
llm_plan: Dict[str, str] = field(default_factory=lambda: {})
|
||||
llm_nickname_mapping: Dict[str, Any] = field(default_factory=dict)
|
||||
llm_nickname_mapping: Dict[str, str] = field(default_factory=lambda: {})
|
||||
llm_scheduler_all: Dict[str, str] = field(default_factory=lambda: {})
|
||||
llm_scheduler_doing: Dict[str, str] = field(default_factory=lambda: {})
|
||||
|
||||
api_urls: Dict[str, str] = field(default_factory=lambda: {})
|
||||
|
||||
|
|
@ -365,9 +369,10 @@ class BotConfig:
|
|||
|
||||
def personality(parent: dict):
|
||||
personality_config = parent["personality"]
|
||||
if config.INNER_VERSION in SpecifierSet(">=1.2.4"):
|
||||
if config.INNER_VERSION in SpecifierSet(">=1.6.1.2"):
|
||||
config.personality_core = personality_config.get("personality_core", config.personality_core)
|
||||
config.personality_sides = personality_config.get("personality_sides", config.personality_sides)
|
||||
config.personality_detail_level = personality_config.get("personality_detail_level", config.personality_sides)
|
||||
|
||||
def identity(parent: dict):
|
||||
identity_config = parent["identity"]
|
||||
|
|
@ -411,7 +416,7 @@ class BotConfig:
|
|||
config.steal_emoji = emoji_config.get("steal_emoji", config.steal_emoji)
|
||||
|
||||
def group_nickname(parent: dict):
|
||||
if config.INNER_VERSION in SpecifierSet(">=1.6.2"):
|
||||
if config.INNER_VERSION in SpecifierSet(">=1.6.1.1"):
|
||||
gn_config = parent.get("group_nickname", {})
|
||||
config.ENABLE_NICKNAME_MAPPING = gn_config.get(
|
||||
"enable_nickname_mapping", config.ENABLE_NICKNAME_MAPPING
|
||||
|
|
@ -515,6 +520,8 @@ class BotConfig:
|
|||
"llm_PFC_chat",
|
||||
"llm_PFC_reply_checker",
|
||||
"llm_nickname_mapping",
|
||||
"llm_scheduler_all",
|
||||
"llm_scheduler_doing",
|
||||
]
|
||||
|
||||
for item in config_list:
|
||||
|
|
@ -720,7 +727,7 @@ class BotConfig:
|
|||
"chat": {"func": chat, "support": ">=1.6.0", "necessary": False},
|
||||
"normal_chat": {"func": normal_chat, "support": ">=1.6.0", "necessary": False},
|
||||
"focus_chat": {"func": focus_chat, "support": ">=1.6.0", "necessary": False},
|
||||
"group_nickname": {"func": group_nickname, "support": ">=0.6.3", "necessary": False},
|
||||
"group_nickname": {"func": group_nickname, "support": ">=1.6.1.1", "necessary": False},
|
||||
}
|
||||
|
||||
# 原地修改，将 字符串版本表达式 转换成 版本对象
|
||||
|
|
|
|||
|
|
@ -413,77 +413,78 @@ class SubMind:
|
|||
logger.warning(f"{self.log_prefix} LLM返回空结果，思考失败。")
|
||||
|
||||
# ---------- 6. 应用概率性去重和修饰 ----------
|
||||
new_content = content # 保存 LLM 直接输出的结果
|
||||
try:
|
||||
similarity = calculate_similarity(previous_mind, new_content)
|
||||
replacement_prob = calculate_replacement_probability(similarity)
|
||||
logger.debug(f"{self.log_prefix} 新旧想法相似度: {similarity:.2f}, 替换概率: {replacement_prob:.2f}")
|
||||
if global_config.allow_remove_duplicates:
|
||||
new_content = content # 保存 LLM 直接输出的结果
|
||||
try:
|
||||
similarity = calculate_similarity(previous_mind, new_content)
|
||||
replacement_prob = calculate_replacement_probability(similarity)
|
||||
logger.debug(f"{self.log_prefix} 新旧想法相似度: {similarity:.2f}, 替换概率: {replacement_prob:.2f}")
|
||||
|
||||
# 定义词语列表 (移到判断之前)
|
||||
yu_qi_ci_liebiao = ["嗯", "哦", "啊", "唉", "哈", "唔"]
|
||||
zhuan_zhe_liebiao = ["但是", "不过", "然而", "可是", "只是"]
|
||||
cheng_jie_liebiao = ["然后", "接着", "此外", "而且", "另外"]
|
||||
zhuan_jie_ci_liebiao = zhuan_zhe_liebiao + cheng_jie_liebiao
|
||||
# 定义词语列表 (移到判断之前)
|
||||
yu_qi_ci_liebiao = ["嗯", "哦", "啊", "唉", "哈", "唔"]
|
||||
zhuan_zhe_liebiao = ["但是", "不过", "然而", "可是", "只是"]
|
||||
cheng_jie_liebiao = ["然后", "接着", "此外", "而且", "另外"]
|
||||
zhuan_jie_ci_liebiao = zhuan_zhe_liebiao + cheng_jie_liebiao
|
||||
|
||||
if random.random() < replacement_prob:
|
||||
# 相似度非常高时，尝试去重或特殊处理
|
||||
if similarity == 1.0:
|
||||
logger.debug(f"{self.log_prefix} 想法完全重复 (相似度 1.0)，执行特殊处理...")
|
||||
# 随机截取大约一半内容
|
||||
if len(new_content) > 1: # 避免内容过短无法截取
|
||||
split_point = max(
|
||||
1, len(new_content) // 2 + random.randint(-len(new_content) // 4, len(new_content) // 4)
|
||||
)
|
||||
truncated_content = new_content[:split_point]
|
||||
else:
|
||||
truncated_content = new_content # 如果只有一个字符或者为空，就不截取了
|
||||
|
||||
# 添加语气词和转折/承接词
|
||||
yu_qi_ci = random.choice(yu_qi_ci_liebiao)
|
||||
zhuan_jie_ci = random.choice(zhuan_jie_ci_liebiao)
|
||||
content = f"{yu_qi_ci}{zhuan_jie_ci}，{truncated_content}"
|
||||
logger.debug(f"{self.log_prefix} 想法重复，特殊处理后: {content}")
|
||||
|
||||
else:
|
||||
# 相似度较高但非100%，执行标准去重逻辑
|
||||
logger.debug(f"{self.log_prefix} 执行概率性去重 (概率: {replacement_prob:.2f})...")
|
||||
matcher = difflib.SequenceMatcher(None, previous_mind, new_content)
|
||||
deduplicated_parts = []
|
||||
last_match_end_in_b = 0
|
||||
for _i, j, n in matcher.get_matching_blocks():
|
||||
if last_match_end_in_b < j:
|
||||
deduplicated_parts.append(new_content[last_match_end_in_b:j])
|
||||
last_match_end_in_b = j + n
|
||||
|
||||
deduplicated_content = "".join(deduplicated_parts).strip()
|
||||
|
||||
if deduplicated_content:
|
||||
# 根据概率决定是否添加词语
|
||||
prefix_str = ""
|
||||
if random.random() < 0.3: # 30% 概率添加语气词
|
||||
prefix_str += random.choice(yu_qi_ci_liebiao)
|
||||
if random.random() < 0.7: # 70% 概率添加转折/承接词
|
||||
prefix_str += random.choice(zhuan_jie_ci_liebiao)
|
||||
|
||||
# 组合最终结果
|
||||
if prefix_str:
|
||||
content = f"{prefix_str}，{deduplicated_content}" # 更新 content
|
||||
logger.debug(f"{self.log_prefix} 去重并添加引导词后: {content}")
|
||||
if random.random() < replacement_prob:
|
||||
# 相似度非常高时，尝试去重或特殊处理
|
||||
if similarity == 1.0:
|
||||
logger.debug(f"{self.log_prefix} 想法完全重复 (相似度 1.0)，执行特殊处理...")
|
||||
# 随机截取大约一半内容
|
||||
if len(new_content) > 1: # 避免内容过短无法截取
|
||||
split_point = max(
|
||||
1, len(new_content) // 2 + random.randint(-len(new_content) // 4, len(new_content) // 4)
|
||||
)
|
||||
truncated_content = new_content[:split_point]
|
||||
else:
|
||||
content = deduplicated_content # 更新 content
|
||||
logger.debug(f"{self.log_prefix} 去重后 (未添加引导词): {content}")
|
||||
else:
|
||||
logger.warning(f"{self.log_prefix} 去重后内容为空，保留原始LLM输出: {new_content}")
|
||||
content = new_content # 保留原始 content
|
||||
else:
|
||||
logger.debug(f"{self.log_prefix} 未执行概率性去重 (概率: {replacement_prob:.2f})")
|
||||
# content 保持 new_content 不变
|
||||
truncated_content = new_content # 如果只有一个字符或者为空，就不截取了
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"{self.log_prefix} 应用概率性去重或特殊处理时出错: {e}")
|
||||
logger.error(traceback.format_exc())
|
||||
# 出错时保留原始 content
|
||||
content = new_content
|
||||
# 添加语气词和转折/承接词
|
||||
yu_qi_ci = random.choice(yu_qi_ci_liebiao)
|
||||
zhuan_jie_ci = random.choice(zhuan_jie_ci_liebiao)
|
||||
content = f"{yu_qi_ci}{zhuan_jie_ci}，{truncated_content}"
|
||||
logger.debug(f"{self.log_prefix} 想法重复，特殊处理后: {content}")
|
||||
|
||||
else:
|
||||
# 相似度较高但非100%，执行标准去重逻辑
|
||||
logger.debug(f"{self.log_prefix} 执行概率性去重 (概率: {replacement_prob:.2f})...")
|
||||
matcher = difflib.SequenceMatcher(None, previous_mind, new_content)
|
||||
deduplicated_parts = []
|
||||
last_match_end_in_b = 0
|
||||
for _i, j, n in matcher.get_matching_blocks():
|
||||
if last_match_end_in_b < j:
|
||||
deduplicated_parts.append(new_content[last_match_end_in_b:j])
|
||||
last_match_end_in_b = j + n
|
||||
|
||||
deduplicated_content = "".join(deduplicated_parts).strip()
|
||||
|
||||
if deduplicated_content:
|
||||
# 根据概率决定是否添加词语
|
||||
prefix_str = ""
|
||||
if random.random() < 0.3: # 30% 概率添加语气词
|
||||
prefix_str += random.choice(yu_qi_ci_liebiao)
|
||||
if random.random() < 0.7: # 70% 概率添加转折/承接词
|
||||
prefix_str += random.choice(zhuan_jie_ci_liebiao)
|
||||
|
||||
# 组合最终结果
|
||||
if prefix_str:
|
||||
content = f"{prefix_str}，{deduplicated_content}" # 更新 content
|
||||
logger.debug(f"{self.log_prefix} 去重并添加引导词后: {content}")
|
||||
else:
|
||||
content = deduplicated_content # 更新 content
|
||||
logger.debug(f"{self.log_prefix} 去重后 (未添加引导词): {content}")
|
||||
else:
|
||||
logger.warning(f"{self.log_prefix} 去重后内容为空，保留原始LLM输出: {new_content}")
|
||||
content = new_content # 保留原始 content
|
||||
else:
|
||||
logger.debug(f"{self.log_prefix} 未执行概率性去重 (概率: {replacement_prob:.2f})")
|
||||
# content 保持 new_content 不变
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"{self.log_prefix} 应用概率性去重或特殊处理时出错: {e}")
|
||||
logger.error(traceback.format_exc())
|
||||
# 出错时保留原始 content
|
||||
content = new_content
|
||||
|
||||
# ---------- 7. 更新思考状态并返回结果 ----------
|
||||
logger.info(f"{self.log_prefix} 最终心流思考结果: {content}")
|
||||
|
|
|
|||
|
|
@ -3,6 +3,7 @@ from .personality import Personality
|
|||
from .identity import Identity
|
||||
import random
|
||||
from rich.traceback import install
|
||||
from src.config.config import global_config
|
||||
|
||||
install(extra_lines=3)
|
||||
|
||||
|
|
@ -205,6 +206,15 @@ class Individuality:
|
|||
if not self.personality or not self.identity:
|
||||
return "个体特征尚未完全初始化。"
|
||||
|
||||
if global_config.personality_detail_level == 1:
|
||||
level = 1
|
||||
elif global_config.personality_detail_level == 2:
|
||||
level = 2
|
||||
elif global_config.personality_detail_level == 3:
|
||||
level = 3
|
||||
else: # level = 0
|
||||
pass
|
||||
|
||||
# 调用新的独立方法
|
||||
prompt_personality = self.get_personality_prompt(level, x_person)
|
||||
prompt_identity = self.get_identity_prompt(level, x_person)
|
||||
|
|
|
|||
|
|
@ -31,13 +31,13 @@ class ScheduleGenerator:
|
|||
# 使用离线LLM模型
|
||||
self.enable_output = None
|
||||
self.llm_scheduler_all = LLMRequest(
|
||||
model=global_config.llm_reasoning,
|
||||
model=global_config.llm_scheduler_all,
|
||||
temperature=global_config.SCHEDULE_TEMPERATURE + 0.3,
|
||||
max_tokens=7000,
|
||||
request_type="schedule",
|
||||
)
|
||||
self.llm_scheduler_doing = LLMRequest(
|
||||
model=global_config.llm_normal,
|
||||
model=global_config.llm_scheduler_all,
|
||||
temperature=global_config.SCHEDULE_TEMPERATURE,
|
||||
max_tokens=2048,
|
||||
request_type="schedule",
|
||||
|
|
|
|||
|
|
@ -1,5 +1,5 @@
|
|||
[inner]
|
||||
version = "1.6.2"
|
||||
version = "1.6.1.2"
|
||||
|
||||
#----以下是给开发人员阅读的，如果你只是部署了麦麦，不需要阅读----
|
||||
#如果你想要修改配置文件，请在修改后将version的值进行变更
|
||||
|
|
@ -42,6 +42,7 @@ personality_sides = [
|
|||
"用一句话或几句话描述人格的一些细节",
|
||||
"用一句话或几句话描述人格的一些细节",
|
||||
]# 条数任意，不能为0, 该选项还在调试中，可能未完全生效
|
||||
personality_detail_level = 0 # 人设消息注入 prompt 详细等级 (0: 采用默认配置, 1: 核心/随机细节, 2: 核心+随机侧面/全部细节, 3: 全部)
|
||||
|
||||
[identity] #アイデンティティがない 生まれないらららら
|
||||
# 兴趣爱好 未完善，有些条目未使用
|
||||
|
|
@ -71,6 +72,7 @@ allow_focus_mode = true # 是否允许专注聊天状态
|
|||
# 启用后麦麦会自主选择进入heart_flowC模式(持续一段时间），进行主动的观察和回复，并给出回复，比较消耗token
|
||||
base_normal_chat_num = 3 # 最多允许多少个群进行普通聊天
|
||||
base_focused_chat_num = 2 # 最多允许多少个群进行专注聊天
|
||||
allow_remove_duplicates = true # 是否开启心流去重（如果发现心流截断问题严重可尝试关闭）
|
||||
|
||||
observation_context_size = 15 # 观察到的最长上下文大小,建议15，太短太长都会导致脑袋尖尖
|
||||
message_buffer = true # 启用消息缓冲器？启用此项以解决消息的拆分问题，但会使麦麦的回复延迟
|
||||
|
|
@ -306,6 +308,22 @@ temp = 0.3
|
|||
pri_in = 2
|
||||
pri_out = 8
|
||||
|
||||
#日程模型
|
||||
[model.llm_scheduler_all]
|
||||
name = "deepseek-ai/DeepSeek-V3"
|
||||
provider = "SILICONFLOW"
|
||||
temp = 0.3
|
||||
pri_in = 2
|
||||
pri_out = 8
|
||||
|
||||
#在干嘛模型
|
||||
[model.llm_scheduler_doing]
|
||||
name = "deepseek-ai/DeepSeek-V3"
|
||||
provider = "SILICONFLOW"
|
||||
temp = 0.3
|
||||
pri_in = 2
|
||||
pri_out = 8
|
||||
|
||||
#以下模型暂时没有使用！！
|
||||
#以下模型暂时没有使用！！
|
||||
#以下模型暂时没有使用！！
|
||||
|
|
|
|||
Loading…
Reference in New Issue