207 lines
8.7 KiB
Python
207 lines
8.7 KiB
Python
from typing import AsyncGenerator
|
||
from src.llm_models.utils_model import LLMRequest, RequestType
|
||
from src.llm_models.payload_content.message import MessageBuilder
|
||
from src.config.config import model_config
|
||
from src.chat.message_receive.message import MessageRecvS4U
|
||
from src.mais4u.mais4u_chat.s4u_prompt import prompt_builder
|
||
from src.common.logger import get_logger
|
||
import re
|
||
|
||
|
||
logger = get_logger("s4u_stream_generator")
|
||
|
||
|
||
class S4UStreamGenerator:
|
||
def __init__(self):
|
||
# 使用LLMRequest替代AsyncOpenAIClient
|
||
self.llm_request = LLMRequest(
|
||
model_set=model_config.model_task_config.replyer,
|
||
request_type="s4u_replyer"
|
||
)
|
||
|
||
self.current_model_name = "unknown model"
|
||
self.partial_response = ""
|
||
|
||
# 正则表达式用于按句子切分，同时处理各种标点和边缘情况
|
||
# 匹配常见的句子结束符，但会忽略引号内和数字中的标点
|
||
self.sentence_split_pattern = re.compile(
|
||
r'([^\s\w"\'([{]*["\'([{].*?["\'}\])][^\s\w"\'([{]*|' # 匹配被引号/括号包裹的内容
|
||
r'[^.。!?？！\n\r]+(?:[.。!?？！\n\r](?![\'"])|$))', # 匹配直到句子结束符
|
||
re.UNICODE | re.DOTALL,
|
||
)
|
||
|
||
self.chat_stream = None
|
||
|
||
async def build_last_internal_message(self, message: MessageRecvS4U, previous_reply_context: str = ""):
|
||
# person_id = PersonInfoManager.get_person_id(
|
||
# message.chat_stream.user_info.platform, message.chat_stream.user_info.user_id
|
||
# )
|
||
# person_info_manager = get_person_info_manager()
|
||
# person_name = await person_info_manager.get_value(person_id, "person_name")
|
||
|
||
# if message.chat_stream.user_info.user_nickname:
|
||
# if person_name:
|
||
# sender_name = f"[{message.chat_stream.user_info.user_nickname}]（你叫ta{person_name}）"
|
||
# else:
|
||
# sender_name = f"[{message.chat_stream.user_info.user_nickname}]"
|
||
# else:
|
||
# sender_name = f"用户({message.chat_stream.user_info.user_id})"
|
||
|
||
# 构建prompt
|
||
if previous_reply_context:
|
||
message_txt = f"""
|
||
你正在回复用户的消息，但中途被打断了。这是已有的对话上下文:
|
||
[你已经对上一条消息说的话]: {previous_reply_context}
|
||
---
|
||
[这是用户发来的新消息, 你需要结合上下文，对此进行回复]:
|
||
{message.processed_plain_text}
|
||
"""
|
||
return True, message_txt
|
||
else:
|
||
message_txt = message.processed_plain_text
|
||
return False, message_txt
|
||
|
||
async def generate_response(
|
||
self, message: MessageRecvS4U, previous_reply_context: str = ""
|
||
) -> AsyncGenerator[str, None]:
|
||
"""根据当前模型类型选择对应的生成函数"""
|
||
# 从global_config中获取模型概率值并选择模型
|
||
self.partial_response = ""
|
||
message_txt = message.processed_plain_text
|
||
if not message.is_internal:
|
||
interupted, message_txt_added = await self.build_last_internal_message(message, previous_reply_context)
|
||
if interupted:
|
||
message_txt = message_txt_added
|
||
|
||
message.chat_stream = self.chat_stream
|
||
prompt = await prompt_builder.build_prompt_normal(
|
||
message=message,
|
||
message_txt=message_txt,
|
||
)
|
||
|
||
logger.info(
|
||
f"{self.current_model_name}思考:{message_txt[:30] + '...' if len(message_txt) > 30 else message_txt}"
|
||
) # noqa: E501
|
||
|
||
# 使用LLMRequest进行流式生成
|
||
async for chunk in self._generate_response_with_llm_request(prompt):
|
||
yield chunk
|
||
|
||
async def _generate_response_with_llm_request(self, prompt: str) -> AsyncGenerator[str, None]:
|
||
"""使用LLMRequest进行流式响应生成"""
|
||
|
||
# 构建消息
|
||
message_builder = MessageBuilder()
|
||
message_builder.add_text_content(prompt)
|
||
messages = [message_builder.build()]
|
||
|
||
# 选择模型
|
||
model_info, api_provider, client = self.llm_request._select_model()
|
||
self.current_model_name = model_info.name
|
||
|
||
# 如果模型支持强制流式模式，使用真正的流式处理
|
||
if model_info.force_stream_mode:
|
||
# 简化流式处理：直接使用LLMRequest的流式功能
|
||
try:
|
||
# 直接调用LLMRequest的流式处理
|
||
response = await self.llm_request._execute_request(
|
||
api_provider=api_provider,
|
||
client=client,
|
||
request_type=RequestType.RESPONSE,
|
||
model_info=model_info,
|
||
message_list=messages,
|
||
)
|
||
|
||
# 处理响应内容
|
||
content = response.content or ""
|
||
if content:
|
||
# 将内容按句子分割并输出
|
||
async for chunk in self._process_content_streaming(content):
|
||
yield chunk
|
||
|
||
except Exception as e:
|
||
logger.error(f"流式请求执行失败: {e}")
|
||
# 如果流式请求失败，回退到普通模式
|
||
response = await self.llm_request._execute_request(
|
||
api_provider=api_provider,
|
||
client=client,
|
||
request_type=RequestType.RESPONSE,
|
||
model_info=model_info,
|
||
message_list=messages,
|
||
)
|
||
content = response.content or ""
|
||
async for chunk in self._process_content_streaming(content):
|
||
yield chunk
|
||
|
||
else:
|
||
# 如果不支持流式，使用普通方式然后模拟流式输出
|
||
response = await self.llm_request._execute_request(
|
||
api_provider=api_provider,
|
||
client=client,
|
||
request_type=RequestType.RESPONSE,
|
||
model_info=model_info,
|
||
message_list=messages,
|
||
)
|
||
|
||
content = response.content or ""
|
||
async for chunk in self._process_content_streaming(content):
|
||
yield chunk
|
||
|
||
async def _process_buffer_streaming(self, buffer: str) -> AsyncGenerator[str, None]:
|
||
"""实时处理缓冲区内容，输出完整句子"""
|
||
# 使用正则表达式匹配完整句子
|
||
for match in self.sentence_split_pattern.finditer(buffer):
|
||
sentence = match.group(0).strip()
|
||
if sentence and match.end(0) <= len(buffer):
|
||
# 检查句子是否完整（以标点符号结尾）
|
||
if sentence.endswith(("。", "！", "？", ".", "!", "?")):
|
||
if sentence not in [",", "，", ".", "。", "!", "！", "?", "？"]:
|
||
self.partial_response += sentence
|
||
yield sentence
|
||
|
||
async def _process_content_streaming(self, content: str) -> AsyncGenerator[str, None]:
|
||
"""处理内容进行流式输出（用于非流式模型的模拟流式输出）"""
|
||
buffer = content
|
||
punctuation_buffer = ""
|
||
|
||
# 使用正则表达式匹配句子
|
||
last_match_end = 0
|
||
for match in self.sentence_split_pattern.finditer(buffer):
|
||
sentence = match.group(0).strip()
|
||
if sentence:
|
||
# 检查是否只是一个标点符号
|
||
if sentence in [",", "，", ".", "。", "!", "！", "?", "？"]:
|
||
punctuation_buffer += sentence
|
||
else:
|
||
# 发送之前累积的标点和当前句子
|
||
to_yield = punctuation_buffer + sentence
|
||
if to_yield.endswith((",", "，")):
|
||
to_yield = to_yield.rstrip(",，")
|
||
|
||
self.partial_response += to_yield
|
||
yield to_yield
|
||
punctuation_buffer = "" # 清空标点符号缓冲区
|
||
|
||
last_match_end = match.end(0)
|
||
|
||
# 发送缓冲区中剩余的任何内容
|
||
remaining = buffer[last_match_end:].strip()
|
||
to_yield = (punctuation_buffer + remaining).strip()
|
||
if to_yield:
|
||
if to_yield.endswith(("，", ",")):
|
||
to_yield = to_yield.rstrip("，,")
|
||
if to_yield:
|
||
self.partial_response += to_yield
|
||
yield to_yield
|
||
|
||
async def _generate_response_with_model(
|
||
self,
|
||
prompt: str,
|
||
client,
|
||
model_name: str,
|
||
**kwargs,
|
||
) -> AsyncGenerator[str, None]:
|
||
"""保留原有方法签名以保持兼容性，但重定向到新的实现"""
|
||
async for chunk in self._generate_response_with_llm_request(prompt):
|
||
yield chunk
|