476 lines
21 KiB
Python
476 lines
21 KiB
Python
from .observation import Observation
|
||
from src.plugins.models.utils_model import LLMRequest
|
||
from src.config.config import global_config
|
||
import time
|
||
import traceback
|
||
from src.common.logger_manager import get_logger
|
||
from src.individuality.individuality import Individuality
|
||
import random
|
||
from ..plugins.utils.prompt_builder import Prompt, global_prompt_manager
|
||
from src.do_tool.tool_use import ToolUser
|
||
from src.plugins.utils.json_utils import safe_json_dumps, process_llm_tool_calls
|
||
from src.heart_flow.chat_state_info import ChatStateInfo
|
||
from src.plugins.chat.chat_stream import chat_manager
|
||
from src.plugins.heartFC_chat.heartFC_Cycleinfo import CycleInfo
|
||
import difflib
|
||
from src.plugins.person_info.relationship_manager import relationship_manager
|
||
|
||
|
||
logger = get_logger("sub_heartflow")
|
||
|
||
|
||
def init_prompt():
|
||
# --- Group Chat Prompt ---
|
||
group_prompt = """
|
||
{extra_info}
|
||
{relation_prompt}
|
||
你的名字是{bot_name},{prompt_personality}
|
||
{last_loop_prompt}
|
||
{cycle_info_block}
|
||
现在是{time_now}，你正在上网，和qq群里的网友们聊天，以下是正在进行的聊天内容：
|
||
{chat_observe_info}
|
||
|
||
你现在{mood_info}
|
||
请仔细阅读当前群聊内容，分析讨论话题和群成员关系，分析你刚刚发言和别人对你的发言的反应，思考你要不要回复。然后思考你是否需要使用函数工具。
|
||
思考并输出你的内心想法
|
||
输出要求：
|
||
1. 根据聊天内容生成你的想法，{hf_do_next}
|
||
2. 不要分点、不要使用表情符号
|
||
3. 避免多余符号(冒号、引号、括号等)
|
||
4. 语言简洁自然，不要浮夸
|
||
5. 如果你刚发言，并且没有人回复你，不要回复
|
||
工具使用说明：
|
||
1. 输出想法后考虑是否需要使用工具
|
||
2. 工具可获取信息或执行操作
|
||
3. 如需处理消息或回复，请使用工具。"""
|
||
Prompt(group_prompt, "sub_heartflow_prompt_before")
|
||
|
||
# --- Private Chat Prompt ---
|
||
private_prompt = """
|
||
{extra_info}
|
||
{relation_prompt}
|
||
你的名字是{bot_name},{prompt_personality}
|
||
{last_loop_prompt}
|
||
{cycle_info_block}
|
||
现在是{time_now}，你正在上网，和 {chat_target_name} 私聊，以下是你们的聊天内容：
|
||
{chat_observe_info}
|
||
|
||
你现在{mood_info}
|
||
请仔细阅读聊天内容，分析你和 {chat_target_name} 的关系，分析你刚刚发言和对方的反应，思考你要不要回复。然后思考你是否需要使用函数工具。
|
||
思考并输出你的内心想法
|
||
输出要求：
|
||
1. 根据聊天内容生成你的想法，{hf_do_next}
|
||
2. 不要分点、不要使用表情符号
|
||
3. 避免多余符号(冒号、引号、括号等)
|
||
4. 语言简洁自然，不要浮夸
|
||
5. 如果你刚发言，对方没有回复你，请谨慎回复
|
||
工具使用说明：
|
||
1. 输出想法后考虑是否需要使用工具
|
||
2. 工具可获取信息或执行操作
|
||
3. 如需处理消息或回复，请使用工具。"""
|
||
Prompt(private_prompt, "sub_heartflow_prompt_private_before") # New template name
|
||
|
||
# --- Last Loop Prompt (remains the same) ---
|
||
last_loop_t = """
|
||
刚刚你的内心想法是：{current_thinking_info}
|
||
{if_replan_prompt}
|
||
"""
|
||
Prompt(last_loop_t, "last_loop")
|
||
|
||
|
||
def calculate_similarity(text_a: str, text_b: str) -> float:
|
||
"""
|
||
计算两个文本字符串的相似度。
|
||
"""
|
||
if not text_a or not text_b:
|
||
return 0.0
|
||
matcher = difflib.SequenceMatcher(None, text_a, text_b)
|
||
return matcher.ratio()
|
||
|
||
|
||
def calculate_replacement_probability(similarity: float) -> float:
|
||
"""
|
||
根据相似度计算替换的概率。
|
||
规则：
|
||
- 相似度 <= 0.4: 概率 = 0
|
||
- 相似度 >= 0.9: 概率 = 1
|
||
- 相似度 == 0.6: 概率 = 0.7
|
||
- 0.4 < 相似度 <= 0.6: 线性插值 (0.4, 0) 到 (0.6, 0.7)
|
||
- 0.6 < 相似度 < 0.9: 线性插值 (0.6, 0.7) 到 (0.9, 1.0)
|
||
"""
|
||
if similarity <= 0.4:
|
||
return 0.0
|
||
elif similarity >= 0.9:
|
||
return 1.0
|
||
elif 0.4 < similarity <= 0.6:
|
||
# p = 3.5 * s - 1.4
|
||
probability = 3.5 * similarity - 1.4
|
||
return max(0.0, probability)
|
||
else: # 0.6 < similarity < 0.9
|
||
# p = s + 0.1
|
||
probability = similarity + 0.1
|
||
return min(1.0, max(0.0, probability))
|
||
|
||
|
||
class SubMind:
|
||
def __init__(self, subheartflow_id: str, chat_state: ChatStateInfo, observations: Observation):
|
||
self.last_active_time = None
|
||
self.subheartflow_id = subheartflow_id
|
||
|
||
self.llm_model = LLMRequest(
|
||
model=global_config.llm_sub_heartflow,
|
||
temperature=global_config.llm_sub_heartflow["temp"],
|
||
max_tokens=800,
|
||
request_type="sub_heart_flow",
|
||
)
|
||
|
||
self.chat_state = chat_state
|
||
self.observations = observations
|
||
|
||
self.current_mind = ""
|
||
self.past_mind = []
|
||
self.structured_info = {}
|
||
|
||
name = chat_manager.get_stream_name(self.subheartflow_id)
|
||
self.log_prefix = f"[{name}] "
|
||
|
||
async def do_thinking_before_reply(self, history_cycle: list[CycleInfo] = None):
|
||
"""
|
||
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
|
||
previous_mind = self.current_mind if self.current_mind else ""
|
||
mood_info = self.chat_state.mood
|
||
|
||
# 获取观察对象
|
||
observation = self.observations[0] if self.observations else None
|
||
if not observation or not hasattr(observation, 'is_group_chat'): # Ensure it's ChattingObservation or similar
|
||
logger.error(f"{self.log_prefix} 无法获取有效的观察对象或缺少聊天类型信息")
|
||
self.update_current_mind("(观察出错了...)")
|
||
return self.current_mind, self.past_mind
|
||
|
||
is_group_chat = observation.is_group_chat
|
||
chat_target_info = observation.chat_target_info
|
||
chat_target_name = "对方" # Default for private
|
||
if not is_group_chat and chat_target_info:
|
||
chat_target_name = chat_target_info.get('person_name') or chat_target_info.get('user_nickname') or chat_target_name
|
||
# --- End getting observation info ---
|
||
|
||
# 获取观察内容
|
||
chat_observe_info = observation.get_observe_info()
|
||
person_list = observation.person_list
|
||
|
||
# ---------- 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()
|
||
|
||
relation_prompt = ""
|
||
# print(f"person_list: {person_list}")
|
||
for person in person_list:
|
||
relation_prompt += await relationship_manager.build_relationship_info(person, is_id=True)
|
||
|
||
# print(f"relat22222ion_prompt: {relation_prompt}")
|
||
|
||
# 构建个性部分
|
||
prompt_personality = individuality.get_prompt(x_person=2, level=2)
|
||
|
||
# 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 = [
|
||
("可以参考之前的想法，在原来想法的基础上继续思考", 0.2),
|
||
("可以参考之前的想法，在原来的想法上尝试新的话题", 0.4),
|
||
("不要太深入", 0.2),
|
||
("进行深入思考", 0.2),
|
||
]
|
||
|
||
last_cycle = history_cycle[-1] if history_cycle else None
|
||
# 上一次决策信息
|
||
if last_cycle is not None:
|
||
last_action = last_cycle.action_type
|
||
last_reasoning = last_cycle.reasoning
|
||
is_replan = last_cycle.replanned
|
||
if is_replan:
|
||
if_replan_prompt = f"但是你有了上述想法之后，有了新消息，你决定重新思考后，你做了：{last_action}\n因为：{last_reasoning}\n"
|
||
else:
|
||
if_replan_prompt = f"出于这个想法，你刚才做了：{last_action}\n因为：{last_reasoning}\n"
|
||
else:
|
||
last_action = ""
|
||
last_reasoning = ""
|
||
is_replan = False
|
||
if_replan_prompt = ""
|
||
if previous_mind:
|
||
last_loop_prompt = (await global_prompt_manager.get_prompt_async("last_loop")).format(
|
||
current_thinking_info=previous_mind, if_replan_prompt=if_replan_prompt
|
||
)
|
||
else:
|
||
last_loop_prompt = ""
|
||
|
||
# 准备循环信息块 (分析最近的活动循环)
|
||
recent_active_cycles = []
|
||
for cycle in reversed(history_cycle):
|
||
# 只关心实际执行了动作的循环
|
||
if cycle.action_taken:
|
||
recent_active_cycles.append(cycle)
|
||
# 最多找最近的3个活动循环
|
||
if len(recent_active_cycles) == 3:
|
||
break
|
||
|
||
cycle_info_block = ""
|
||
consecutive_text_replies = 0
|
||
responses_for_prompt = []
|
||
|
||
# 检查这最近的活动循环中有多少是连续的文本回复 (从最近的开始看)
|
||
for cycle in recent_active_cycles:
|
||
if cycle.action_type == "text_reply":
|
||
consecutive_text_replies += 1
|
||
# 获取回复内容，如果不存在则返回'[空回复]'
|
||
response_text = cycle.response_info.get("response_text", [])
|
||
# 使用简单的 join 来格式化回复内容列表
|
||
formatted_response = "[空回复]" if not response_text else " ".join(response_text)
|
||
responses_for_prompt.append(formatted_response)
|
||
else:
|
||
# 一旦遇到非文本回复，连续性中断
|
||
break
|
||
|
||
# 根据连续文本回复的数量构建提示信息
|
||
# 注意: responses_for_prompt 列表是从最近到最远排序的
|
||
if consecutive_text_replies >= 3: # 如果最近的三个活动都是文本回复
|
||
cycle_info_block = f'你已经连续回复了三条消息（最近: "{responses_for_prompt[0]}"，第二近: "{responses_for_prompt[1]}"，第三近: "{responses_for_prompt[2]}"）。你回复的有点多了，请注意'
|
||
elif consecutive_text_replies == 2: # 如果最近的两个活动是文本回复
|
||
cycle_info_block = f'你已经连续回复了两条消息（最近: "{responses_for_prompt[0]}"，第二近: "{responses_for_prompt[1]}"），请注意'
|
||
elif consecutive_text_replies == 1: # 如果最近的一个活动是文本回复
|
||
cycle_info_block = f'你刚刚已经回复一条消息（内容: "{responses_for_prompt[0]}"）'
|
||
|
||
# 包装提示块，增加可读性，即使没有连续回复也给个标记
|
||
if cycle_info_block:
|
||
cycle_info_block = f"\n【近期回复历史】\n{cycle_info_block}\n"
|
||
else:
|
||
# 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"
|
||
|
||
# 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
|
||
)[0]
|
||
|
||
# ---------- 4. 构建最终提示词 ----------
|
||
# --- Choose template based on chat type ---
|
||
if is_group_chat:
|
||
template_name = "sub_heartflow_prompt_before"
|
||
prompt = (await global_prompt_manager.get_prompt_async(template_name)).format(
|
||
extra_info="",
|
||
prompt_personality=prompt_personality,
|
||
relation_prompt=relation_prompt,
|
||
bot_name=individuality.name,
|
||
time_now=time_now,
|
||
chat_observe_info=chat_observe_info,
|
||
mood_info=mood_info,
|
||
hf_do_next=hf_do_next,
|
||
last_loop_prompt=last_loop_prompt,
|
||
cycle_info_block=cycle_info_block,
|
||
# chat_target_name is not used in group prompt
|
||
)
|
||
else: # Private chat
|
||
template_name = "sub_heartflow_prompt_private_before"
|
||
prompt = (await global_prompt_manager.get_prompt_async(template_name)).format(
|
||
extra_info="",
|
||
prompt_personality=prompt_personality,
|
||
relation_prompt=relation_prompt, # Might need adjustment for private context
|
||
bot_name=individuality.name,
|
||
time_now=time_now,
|
||
chat_target_name=chat_target_name, # Pass target name
|
||
chat_observe_info=chat_observe_info,
|
||
mood_info=mood_info,
|
||
hf_do_next=hf_do_next,
|
||
last_loop_prompt=last_loop_prompt,
|
||
cycle_info_block=cycle_info_block,
|
||
)
|
||
# --- End choosing template ---
|
||
|
||
# ---------- 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:
|
||
# 调用LLM生成响应
|
||
response, _reasoning_content, tool_calls = await self.llm_model.generate_response_tool_async(
|
||
prompt=prompt, tools=tools
|
||
)
|
||
|
||
logger.debug(f"{self.log_prefix} 子心流输出的原始LLM响应: {response}")
|
||
|
||
# 直接使用LLM返回的文本响应作为 content
|
||
content = response if response else ""
|
||
|
||
if tool_calls:
|
||
# 直接将 tool_calls 传递给处理函数
|
||
success, valid_tool_calls, error_msg = process_llm_tool_calls(
|
||
tool_calls, log_prefix=f"{self.log_prefix} "
|
||
)
|
||
|
||
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.log_prefix} 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.log_prefix} 处理工具调用时出错: {error_msg}")
|
||
else:
|
||
logger.info(f"{self.log_prefix} 心流未使用工具")
|
||
|
||
except Exception as e:
|
||
# 处理总体异常
|
||
logger.error(f"{self.log_prefix} 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"
|
||
|
||
# 记录初步思考结果
|
||
logger.debug(f"{self.log_prefix} 初步心流思考结果: {content}\nprompt: {prompt}\n")
|
||
|
||
# 处理空响应情况
|
||
if not content:
|
||
content = "(不知道该想些什么...)"
|
||
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}")
|
||
|
||
# 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}")
|
||
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}")
|
||
# 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)
|
||
self.current_mind = response
|
||
|
||
|
||
init_prompt()
|