import asyncio
import json
import re
from datetime import datetime
from typing import Tuple, Union
import aiohttp
from loguru import logger
from nonebot import get_driver
import base64
from PIL import Image
import io
from ...common.database import db
from ..chat.config import global_config
driver = get_driver()
config = driver.config
class LLM_request:
def __init__(self, model, **kwargs):
# 将大写的配置键转换为小写并从config中获取实际值
try:
self.api_key = getattr(config, model["key"])
self.base_url = getattr(config, model["base_url"])
except AttributeError as e:
logger.error(f"原始 model dict 信息：{model}")
logger.error(f"配置错误：找不到对应的配置项 - {str(e)}")
raise ValueError(f"配置错误：找不到对应的配置项 - {str(e)}") from e
self.model_name = model["name"]
self.params = kwargs
self.pri_in = model.get("pri_in", 0)
self.pri_out = model.get("pri_out", 0)
# 获取数据库实例
self._init_database()
def _init_database(self):
"""初始化数据库集合"""
try:
# 创建llm_usage集合的索引
db.llm_usage.create_index([("timestamp", 1)])
db.llm_usage.create_index([("model_name", 1)])
db.llm_usage.create_index([("user_id", 1)])
db.llm_usage.create_index([("request_type", 1)])
except Exception:
logger.error("创建数据库索引失败")
def _record_usage(self, prompt_tokens: int, completion_tokens: int, total_tokens: int,
user_id: str = "system", request_type: str = "chat",
endpoint: str = "/chat/completions"):
"""记录模型使用情况到数据库
Args:
prompt_tokens: 输入token数
completion_tokens: 输出token数
total_tokens: 总token数
user_id: 用户ID，默认为system
request_type: 请求类型(chat/embedding/image等)
endpoint: API端点
"""
try:
usage_data = {
"model_name": self.model_name,
"user_id": user_id,
"request_type": request_type,
"endpoint": endpoint,
"prompt_tokens": prompt_tokens,
"completion_tokens": completion_tokens,
"total_tokens": total_tokens,
"cost": self._calculate_cost(prompt_tokens, completion_tokens),
"status": "success",
"timestamp": datetime.now()
}
db.llm_usage.insert_one(usage_data)
logger.info(
f"Token使用情况 - 模型: {self.model_name}, "
f"用户: {user_id}, 类型: {request_type}, "
f"提示词: {prompt_tokens}, 完成: {completion_tokens}, "
f"总计: {total_tokens}"
)
except Exception:
logger.error("记录token使用情况失败")
def _calculate_cost(self, prompt_tokens: int, completion_tokens: int) -> float:
"""计算API调用成本
使用模型的pri_in和pri_out价格计算输入和输出的成本
Args:
prompt_tokens: 输入token数量
completion_tokens: 输出token数量
Returns:
float: 总成本（元）
"""
# 使用模型的pri_in和pri_out计算成本
input_cost = (prompt_tokens / 1000000) * self.pri_in
output_cost = (completion_tokens / 1000000) * self.pri_out
return round(input_cost + output_cost, 6)
async def _execute_request(
self,
endpoint: str,
prompt: str = None,
image_base64: str = None,
image_format: str = None,
payload: dict = None,
retry_policy: dict = None,
response_handler: callable = None,
user_id: str = "system",
request_type: str = "chat"
):
"""统一请求执行入口
Args:
endpoint: API端点路径 (如 "chat/completions")
prompt: prompt文本
image_base64: 图片的base64编码
image_format: 图片格式
payload: 请求体数据
retry_policy: 自定义重试策略
response_handler: 自定义响应处理器
user_id: 用户ID
request_type: 请求类型
"""
# 合并重试策略
default_retry = {
"max_retries": 3, "base_wait": 15,
"retry_codes": [429, 413, 500, 503],
"abort_codes": [400, 401, 402, 403]}
policy = {**default_retry, **(retry_policy or {})}
# 常见Error Code Mapping
error_code_mapping = {
400: "参数不正确",
401: "API key 错误，认证失败",
402: "账号余额不足",
403: "需要实名,或余额不足",
404: "Not Found",
429: "请求过于频繁，请稍后再试",
500: "服务器内部故障",
503: "服务器负载过高"
}
api_url = f"{self.base_url.rstrip('/')}/{endpoint.lstrip('/')}"
# 判断是否为流式
stream_mode = self.params.get("stream", False)
if self.params.get("stream", False) is True:
logger.debug(f"进入流式输出模式，发送请求到URL: {api_url}")
else:
logger.debug(f"发送请求到URL: {api_url}")
logger.info(f"使用模型: {self.model_name}")
# 构建请求体
if image_base64:
payload = await self._build_payload(prompt, image_base64, image_format)
elif payload is None:
payload = await self._build_payload(prompt)
for retry in range(policy["max_retries"]):
try:
# 使用上下文管理器处理会话
headers = await self._build_headers()
# 似乎是openai流式必须要的东西,不过阿里云的qwq-plus加了这个没有影响
if stream_mode:
headers["Accept"] = "text/event-stream"
async with aiohttp.ClientSession() as session:
async with session.post(api_url, headers=headers, json=payload) as response:
# 处理需要重试的状态码
if response.status in policy["retry_codes"]:
wait_time = policy["base_wait"] * (2 ** retry)
logger.warning(f"错误码: {response.status}, 等待 {wait_time}秒后重试")
if response.status == 413:
logger.warning("请求体过大，尝试压缩...")
image_base64 = compress_base64_image_by_scale(image_base64)
payload = await self._build_payload(prompt, image_base64, image_format)
elif response.status in [500, 503]:
logger.error(f"错误码: {response.status} - {error_code_mapping.get(response.status)}")
raise RuntimeError("服务器负载过高，模型恢复失败QAQ")
else:
logger.warning(f"请求限制(429)，等待{wait_time}秒后重试...")
await asyncio.sleep(wait_time)
continue
elif response.status in policy["abort_codes"]:
logger.error(f"错误码: {response.status} - {error_code_mapping.get(response.status)}")
if response.status == 403:
#只针对硅基流动的V3和R1进行降级处理
if self.model_name.startswith(
"Pro/deepseek-ai") and self.base_url == "https://api.siliconflow.cn/v1/":
old_model_name = self.model_name
self.model_name = self.model_name[4:] # 移除"Pro/"前缀
logger.warning(f"检测到403错误，模型从 {old_model_name} 降级为 {self.model_name}")
# 对全局配置进行更新
if global_config.llm_normal.get('name') == old_model_name:
global_config.llm_normal['name'] = self.model_name
logger.warning(f"将全局配置中的 llm_normal 模型临时降级至{self.model_name}")
if global_config.llm_reasoning.get('name') == old_model_name:
global_config.llm_reasoning['name'] = self.model_name
logger.warning(f"将全局配置中的 llm_reasoning 模型临时降级至{self.model_name}")
# 更新payload中的模型名
if payload and 'model' in payload:
payload['model'] = self.model_name
# 重新尝试请求
retry -= 1 # 不计入重试次数
continue
raise RuntimeError(f"请求被拒绝: {error_code_mapping.get(response.status)}")
response.raise_for_status()
# 将流式输出转化为非流式输出
if stream_mode:
flag_delta_content_finished = False
accumulated_content = ""
async for line_bytes in response.content:
line = line_bytes.decode("utf-8").strip()
if not line:
continue
if line.startswith("data:"):
data_str = line[5:].strip()
if data_str == "[DONE]":
break
try:
chunk = json.loads(data_str)
if flag_delta_content_finished:
usage = chunk.get("usage", None) # 获取tokn用量
else:
delta = chunk["choices"][0]["delta"]
delta_content = delta.get("content")
if delta_content is None:
delta_content = ""
accumulated_content += delta_content
# 检测流式输出文本是否结束
finish_reason = chunk["choices"][0].get("finish_reason")
if finish_reason == "stop":
usage = chunk.get("usage", None)
if usage:
break
# 部分平台在文本输出结束前不会返回token用量，此时需要再获取一次chunk
flag_delta_content_finished = True
except Exception:
logger.exception("解析流式输出错误")
content = accumulated_content
reasoning_content = ""
think_match = re.search(r'(.*?)', content, re.DOTALL)
if think_match:
reasoning_content = think_match.group(1).strip()
content = re.sub(r'.*?', '', content, flags=re.DOTALL).strip()
# 构造一个伪result以便调用自定义响应处理器或默认处理器
result = {
"choices": [{"message": {"content": content, "reasoning_content": reasoning_content}}], "usage": usage}
return response_handler(result) if response_handler else self._default_response_handler(
result, user_id, request_type, endpoint)
else:
result = await response.json()
# 使用自定义处理器或默认处理
return response_handler(result) if response_handler else self._default_response_handler(
result, user_id, request_type, endpoint)
except Exception as e:
if retry < policy["max_retries"] - 1:
wait_time = policy["base_wait"] * (2 ** retry)
logger.error(f"请求失败，等待{wait_time}秒后重试... 错误: {str(e)}")
await asyncio.sleep(wait_time)
else:
logger.critical(f"请求失败: {str(e)}")
logger.critical(f"请求头: {await self._build_headers(no_key=True)} 请求体: {payload}")
raise RuntimeError(f"API请求失败: {str(e)}")
logger.error("达到最大重试次数，请求仍然失败")
raise RuntimeError("达到最大重试次数，API请求仍然失败")
async def _transform_parameters(self, params: dict) -> dict:
"""
根据模型名称转换参数：
- 对于需要转换的OpenAI CoT系列模型（例如 "o3-mini"），删除 'temprature' 参数，
并将 'max_tokens' 重命名为 'max_completion_tokens'
"""
# 复制一份参数，避免直接修改原始数据
new_params = dict(params)
# 定义需要转换的模型列表
models_needing_transformation = ["o3-mini", "o1-mini", "o1-preview", "o1-2024-12-17", "o1-preview-2024-09-12",
"o3-mini-2025-01-31", "o1-mini-2024-09-12"]
if self.model_name.lower() in models_needing_transformation:
# 删除 'temprature' 参数（如果存在）
new_params.pop("temperature", None)
# 如果存在 'max_tokens'，则重命名为 'max_completion_tokens'
if "max_tokens" in new_params:
new_params["max_completion_tokens"] = new_params.pop("max_tokens")
return new_params
async def _build_payload(self, prompt: str, image_base64: str = None, image_format: str = None) -> dict:
"""构建请求体"""
# 复制一份参数，避免直接修改 self.params
params_copy = await self._transform_parameters(self.params)
if image_base64:
payload = {
"model": self.model_name,
"messages": [
{
"role": "user",
"content": [
{"type": "text", "text": prompt},
{"type": "image_url", "image_url": {"url": f"data:image/{image_format.lower()};base64,{image_base64}"}}
]
}
],
"max_tokens": global_config.max_response_length,
**params_copy
}
else:
payload = {
"model": self.model_name,
"messages": [{"role": "user", "content": prompt}],
"max_tokens": global_config.max_response_length,
**params_copy
}
# 如果 payload 中依然存在 max_tokens 且需要转换，在这里进行再次检查
if self.model_name.lower() in ["o3-mini", "o1-mini", "o1-preview", "o1-2024-12-17", "o1-preview-2024-09-12",
"o3-mini-2025-01-31", "o1-mini-2024-09-12"] and "max_tokens" in payload:
payload["max_completion_tokens"] = payload.pop("max_tokens")
return payload
def _default_response_handler(self, result: dict, user_id: str = "system",
request_type: str = "chat", endpoint: str = "/chat/completions") -> Tuple:
"""默认响应解析"""
if "choices" in result and result["choices"]:
message = result["choices"][0]["message"]
content = message.get("content", "")
content, reasoning = self._extract_reasoning(content)
reasoning_content = message.get("model_extra", {}).get("reasoning_content", "")
if not reasoning_content:
reasoning_content = message.get("reasoning_content", "")
if not reasoning_content:
reasoning_content = reasoning
# 记录token使用情况
usage = result.get("usage", {})
if usage:
prompt_tokens = usage.get("prompt_tokens", 0)
completion_tokens = usage.get("completion_tokens", 0)
total_tokens = usage.get("total_tokens", 0)
self._record_usage(
prompt_tokens=prompt_tokens,
completion_tokens=completion_tokens,
total_tokens=total_tokens,
user_id=user_id,
request_type=request_type,
endpoint=endpoint
)
return content, reasoning_content
return "没有返回结果", ""
def _extract_reasoning(self, content: str) -> tuple[str, str]:
"""CoT思维链提取"""
match = re.search(r'(?:)?(.*?)', content, re.DOTALL)
content = re.sub(r'(?:)?.*?', '', content, flags=re.DOTALL, count=1).strip()
if match:
reasoning = match.group(1).strip()
else:
reasoning = ""
return content, reasoning
async def _build_headers(self, no_key: bool = False) -> dict:
"""构建请求头"""
if no_key:
return {
"Authorization": "Bearer **********",
"Content-Type": "application/json"
}
else:
return {
"Authorization": f"Bearer {self.api_key}",
"Content-Type": "application/json"
}
# 防止小朋友们截图自己的key
async def generate_response(self, prompt: str) -> Tuple[str, str]:
"""根据输入的提示生成模型的异步响应"""
content, reasoning_content = await self._execute_request(
endpoint="/chat/completions",
prompt=prompt
)
return content, reasoning_content
async def generate_response_for_image(self, prompt: str, image_base64: str, image_format: str) -> Tuple[str, str]:
"""根据输入的提示和图片生成模型的异步响应"""
content, reasoning_content = await self._execute_request(
endpoint="/chat/completions",
prompt=prompt,
image_base64=image_base64,
image_format=image_format
)
return content, reasoning_content
async def generate_response_async(self, prompt: str, **kwargs) -> Union[str, Tuple[str, str]]:
"""异步方式根据输入的提示生成模型的响应"""
# 构建请求体
data = {
"model": self.model_name,
"messages": [{"role": "user", "content": prompt}],
"max_tokens": global_config.max_response_length,
**self.params
}
content, reasoning_content = await self._execute_request(
endpoint="/chat/completions",
payload=data,
prompt=prompt
)
return content, reasoning_content
async def get_embedding(self, text: str) -> Union[list, None]:
"""异步方法：获取文本的embedding向量
Args:
text: 需要获取embedding的文本
Returns:
list: embedding向量，如果失败则返回None
"""
def embedding_handler(result):
"""处理响应"""
if "data" in result and len(result["data"]) > 0:
return result["data"][0].get("embedding", None)
return None
embedding = await self._execute_request(
endpoint="/embeddings",
prompt=text,
payload={
"model": self.model_name,
"input": text,
"encoding_format": "float"
},
retry_policy={
"max_retries": 2,
"base_wait": 6
},
response_handler=embedding_handler
)
return embedding
def compress_base64_image_by_scale(base64_data: str, target_size: int = 0.8 * 1024 * 1024) -> str:
"""压缩base64格式的图片到指定大小
Args:
base64_data: base64编码的图片数据
target_size: 目标文件大小（字节），默认0.8MB
Returns:
str: 压缩后的base64图片数据
"""
try:
# 将base64转换为字节数据
image_data = base64.b64decode(base64_data)
# 如果已经小于目标大小，直接返回原图
if len(image_data) <= 2*1024*1024:
return base64_data
# 将字节数据转换为图片对象
img = Image.open(io.BytesIO(image_data))
# 获取原始尺寸
original_width, original_height = img.size
# 计算缩放比例
scale = min(1.0, (target_size / len(image_data)) ** 0.5)
# 计算新的尺寸
new_width = int(original_width * scale)
new_height = int(original_height * scale)
# 创建内存缓冲区
output_buffer = io.BytesIO()
# 如果是GIF，处理所有帧
if getattr(img, "is_animated", False):
frames = []
for frame_idx in range(img.n_frames):
img.seek(frame_idx)
new_frame = img.copy()
new_frame = new_frame.resize((new_width//2, new_height//2), Image.Resampling.LANCZOS) # 动图折上折
frames.append(new_frame)
# 保存到缓冲区
frames[0].save(
output_buffer,
format='GIF',
save_all=True,
append_images=frames[1:],
optimize=True,
duration=img.info.get('duration', 100),
loop=img.info.get('loop', 0)
)
else:
# 处理静态图片
resized_img = img.resize((new_width, new_height), Image.Resampling.LANCZOS)
# 保存到缓冲区，保持原始格式
if img.format == 'PNG' and img.mode in ('RGBA', 'LA'):
resized_img.save(output_buffer, format='PNG', optimize=True)
else:
resized_img.save(output_buffer, format='JPEG', quality=95, optimize=True)
# 获取压缩后的数据并转换为base64
compressed_data = output_buffer.getvalue()
logger.success(f"压缩图片: {original_width}x{original_height} -> {new_width}x{new_height}")
logger.info(f"压缩前大小: {len(image_data)/1024:.1f}KB, 压缩后大小: {len(compressed_data)/1024:.1f}KB")
return base64.b64encode(compressed_data).decode('utf-8')
except Exception as e:
logger.error(f"压缩图片失败: {str(e)}")
import traceback
logger.error(traceback.format_exc())
return base64_data