299 lines
11 KiB
Python
299 lines
11 KiB
Python
import json
|
||
import time
|
||
import random
|
||
import hashlib
|
||
|
||
from typing import List, Dict, Tuple, Optional, Any
|
||
from json_repair import repair_json
|
||
|
||
from src.llm_models.utils_model import LLMRequest
|
||
from src.config.config import global_config, model_config
|
||
from src.common.logger import get_logger
|
||
from src.common.database.database_model import Expression
|
||
from src.chat.utils.prompt_builder import Prompt, global_prompt_manager
|
||
|
||
logger = get_logger("expression_selector")
|
||
|
||
|
||
def init_prompt():
|
||
expression_evaluation_prompt = """
|
||
以下是正在进行的聊天内容：
|
||
{chat_observe_info}
|
||
|
||
你的名字是{bot_name}{target_message}
|
||
|
||
你知道以下这些表达方式，梗和说话方式：
|
||
{all_situations}
|
||
|
||
现在，请你根据聊天记录从中挑选合适的表达方式，梗和说话方式，组织一条回复风格指导，指导的目的是在组织回复的时候提供一些语言风格和梗上的参考。
|
||
请在reply_style_guide中以平文本输出指导，不要浮夸，并在selected_expressions中说明在指导中你挑选了哪些表达方式，梗和说话方式，以json格式输出：
|
||
例子：
|
||
{{
|
||
"reply_style_guide": "...",
|
||
"selected_expressions": [2, 3, 4, 7]
|
||
}}
|
||
请严格按照JSON格式输出，不要包含其他内容：
|
||
"""
|
||
Prompt(expression_evaluation_prompt, "expression_evaluation_prompt")
|
||
|
||
|
||
def weighted_sample(population: List[Dict], weights: List[float], k: int) -> List[Dict]:
|
||
"""按权重随机抽样"""
|
||
if not population or not weights or k <= 0:
|
||
return []
|
||
|
||
if len(population) <= k:
|
||
return population.copy()
|
||
|
||
# 使用累积权重的方法进行加权抽样
|
||
selected = []
|
||
population_copy = population.copy()
|
||
weights_copy = weights.copy()
|
||
|
||
for _ in range(k):
|
||
if not population_copy:
|
||
break
|
||
|
||
# 选择一个元素
|
||
chosen_idx = random.choices(range(len(population_copy)), weights=weights_copy)[0]
|
||
selected.append(population_copy.pop(chosen_idx))
|
||
weights_copy.pop(chosen_idx)
|
||
|
||
return selected
|
||
|
||
|
||
class ExpressionSelector:
|
||
def __init__(self):
|
||
self.llm_model = LLMRequest(
|
||
model_set=model_config.model_task_config.utils_small, request_type="expression.selector"
|
||
)
|
||
|
||
def can_use_expression_for_chat(self, chat_id: str) -> bool:
|
||
"""
|
||
检查指定聊天流是否允许使用表达
|
||
|
||
Args:
|
||
chat_id: 聊天流ID
|
||
|
||
Returns:
|
||
bool: 是否允许使用表达
|
||
"""
|
||
try:
|
||
use_expression, _, _ = global_config.expression.get_expression_config_for_chat(chat_id)
|
||
return use_expression
|
||
except Exception as e:
|
||
logger.error(f"检查表达使用权限失败: {e}")
|
||
return False
|
||
|
||
@staticmethod
|
||
def _parse_stream_config_to_chat_id(stream_config_str: str) -> Optional[str]:
|
||
"""解析'platform:id:type'为chat_id（与get_stream_id一致）"""
|
||
try:
|
||
parts = stream_config_str.split(":")
|
||
if len(parts) != 3:
|
||
return None
|
||
platform = parts[0]
|
||
id_str = parts[1]
|
||
stream_type = parts[2]
|
||
is_group = stream_type == "group"
|
||
if is_group:
|
||
components = [platform, str(id_str)]
|
||
else:
|
||
components = [platform, str(id_str), "private"]
|
||
key = "_".join(components)
|
||
return hashlib.md5(key.encode()).hexdigest()
|
||
except Exception:
|
||
return None
|
||
|
||
def get_related_chat_ids(self, chat_id: str) -> List[str]:
|
||
"""根据expression_groups配置，获取与当前chat_id相关的所有chat_id（包括自身）"""
|
||
groups = global_config.expression.expression_groups
|
||
for group in groups:
|
||
group_chat_ids = []
|
||
for stream_config_str in group:
|
||
if chat_id_candidate := self._parse_stream_config_to_chat_id(stream_config_str):
|
||
group_chat_ids.append(chat_id_candidate)
|
||
if chat_id in group_chat_ids:
|
||
return group_chat_ids
|
||
return [chat_id]
|
||
|
||
def get_random_expressions(
|
||
self, chat_id: str, total_num: int
|
||
) -> List[Dict[str, Any]]:
|
||
# sourcery skip: extract-duplicate-method, move-assign
|
||
# 支持多chat_id合并抽选
|
||
related_chat_ids = self.get_related_chat_ids(chat_id)
|
||
|
||
# 优化：一次性查询所有相关chat_id的表达方式
|
||
style_query = Expression.select().where(
|
||
(Expression.chat_id.in_(related_chat_ids)) & (Expression.type == "style")
|
||
)
|
||
|
||
style_exprs = [
|
||
{
|
||
"situation": expr.situation,
|
||
"style": expr.style,
|
||
"count": expr.count,
|
||
"last_active_time": expr.last_active_time,
|
||
"source_id": expr.chat_id,
|
||
"type": "style",
|
||
"create_date": expr.create_date if expr.create_date is not None else expr.last_active_time,
|
||
}
|
||
for expr in style_query
|
||
]
|
||
|
||
# 按权重抽样（使用count作为权重）
|
||
if style_exprs:
|
||
style_weights = [expr.get("count", 1) for expr in style_exprs]
|
||
selected_style = weighted_sample(style_exprs, style_weights, total_num)
|
||
else:
|
||
selected_style = []
|
||
return selected_style
|
||
|
||
def update_expressions_count_batch(self, expressions_to_update: List[Dict[str, Any]], increment: float = 0.1):
|
||
"""对一批表达方式更新count值，按chat_id+type分组后一次性写入数据库"""
|
||
if not expressions_to_update:
|
||
return
|
||
updates_by_key = {}
|
||
for expr in expressions_to_update:
|
||
source_id: str = expr.get("source_id") # type: ignore
|
||
expr_type: str = expr.get("type", "style")
|
||
situation: str = expr.get("situation") # type: ignore
|
||
style: str = expr.get("style") # type: ignore
|
||
if not source_id or not situation or not style:
|
||
logger.warning(f"表达方式缺少必要字段，无法更新: {expr}")
|
||
continue
|
||
key = (source_id, expr_type, situation, style)
|
||
if key not in updates_by_key:
|
||
updates_by_key[key] = expr
|
||
for chat_id, expr_type, situation, style in updates_by_key:
|
||
query = Expression.select().where(
|
||
(Expression.chat_id == chat_id)
|
||
& (Expression.type == expr_type)
|
||
& (Expression.situation == situation)
|
||
& (Expression.style == style)
|
||
)
|
||
if query.exists():
|
||
expr_obj = query.get()
|
||
current_count = expr_obj.count
|
||
new_count = min(current_count + increment, 5.0)
|
||
expr_obj.count = new_count
|
||
expr_obj.last_active_time = time.time()
|
||
expr_obj.save()
|
||
logger.debug(
|
||
f"表达方式激活: 原count={current_count:.3f}, 增量={increment}, 新count={new_count:.3f} in db"
|
||
)
|
||
|
||
async def select_suitable_expressions_llm(
|
||
self,
|
||
chat_id: str,
|
||
chat_info: str,
|
||
max_num: int = 10,
|
||
target_message: Optional[str] = None,
|
||
) -> Tuple[str, List[Dict[str, Any]]]:
|
||
# sourcery skip: inline-variable, list-comprehension
|
||
"""使用LLM选择适合的表达方式"""
|
||
|
||
# 检查是否允许在此聊天流中使用表达
|
||
if not self.can_use_expression_for_chat(chat_id):
|
||
logger.debug(f"聊天流 {chat_id} 不允许使用表达，返回空列表")
|
||
return "", []
|
||
|
||
# 1. 获取20个随机表达方式（现在按权重抽取）
|
||
style_exprs = self.get_random_expressions(chat_id, 10)
|
||
|
||
# 2. 构建所有表达方式的索引和情境列表
|
||
all_expressions = []
|
||
all_situations = []
|
||
|
||
# 添加style表达方式
|
||
for expr in style_exprs:
|
||
if isinstance(expr, dict) and "situation" in expr and "style" in expr:
|
||
expr_with_type = expr.copy()
|
||
expr_with_type["type"] = "style"
|
||
all_expressions.append(expr_with_type)
|
||
all_situations.append(f"{len(all_expressions)}.当 {expr['situation']} 时，使用 {expr['style']}")
|
||
|
||
if not all_expressions:
|
||
logger.warning("没有找到可用的表达方式")
|
||
return "", []
|
||
|
||
all_situations_str = "\n".join(all_situations)
|
||
|
||
if target_message:
|
||
target_message_str = f"，现在你想要回复消息：{target_message}"
|
||
target_message_extra_block = "4.考虑你要回复的目标消息"
|
||
else:
|
||
target_message_str = ""
|
||
target_message_extra_block = ""
|
||
|
||
# 3. 构建prompt（只包含情境，不包含完整的表达方式）
|
||
prompt = (await global_prompt_manager.get_prompt_async("expression_evaluation_prompt")).format(
|
||
bot_name=global_config.bot.nickname,
|
||
chat_observe_info=chat_info,
|
||
all_situations=all_situations_str,
|
||
max_num=max_num,
|
||
target_message=target_message_str,
|
||
target_message_extra_block=target_message_extra_block,
|
||
)
|
||
|
||
print(prompt)
|
||
|
||
# 4. 调用LLM
|
||
try:
|
||
|
||
# start_time = time.time()
|
||
content, (reasoning_content, model_name, _) = await self.llm_model.generate_response_async(prompt=prompt)
|
||
# logger.info(f"LLM请求时间: {model_name} {time.time() - start_time} \n{prompt}")
|
||
|
||
# logger.info(f"模型名称: {model_name}")
|
||
logger.info(f"LLM返回结果: {content}")
|
||
# if reasoning_content:
|
||
# logger.info(f"LLM推理: {reasoning_content}")
|
||
# else:
|
||
# logger.info(f"LLM推理: 无")
|
||
|
||
if not content:
|
||
logger.warning("LLM返回空结果")
|
||
return "", []
|
||
|
||
# 5. 解析结果
|
||
result = repair_json(content)
|
||
if isinstance(result, str):
|
||
result = json.loads(result)
|
||
|
||
if not isinstance(result, dict) or "reply_style_guide" not in result or "selected_expressions" not in result:
|
||
logger.error("LLM返回格式错误")
|
||
logger.info(f"LLM返回结果: \n{content}")
|
||
return "", []
|
||
|
||
reply_style_guide = result["reply_style_guide"]
|
||
selected_expressions = result["selected_expressions"]
|
||
|
||
# 根据索引获取完整的表达方式
|
||
valid_expressions = []
|
||
for idx in selected_expressions:
|
||
if isinstance(idx, int) and 1 <= idx <= len(all_expressions):
|
||
expression = all_expressions[idx - 1] # 索引从1开始
|
||
valid_expressions.append(expression)
|
||
|
||
# 对选中的所有表达方式，一次性更新count数
|
||
if valid_expressions:
|
||
self.update_expressions_count_batch(valid_expressions, 0.006)
|
||
|
||
# logger.info(f"LLM从{len(all_expressions)}个情境中选择了{len(valid_expressions)}个")
|
||
return reply_style_guide, valid_expressions
|
||
|
||
except Exception as e:
|
||
logger.error(f"LLM处理表达方式选择时出错: {e}")
|
||
return "", []
|
||
|
||
|
||
|
||
init_prompt()
|
||
|
||
try:
|
||
expression_selector = ExpressionSelector()
|
||
except Exception as e:
|
||
print(f"ExpressionSelector初始化失败: {e}")
|