api/index.js

import grpc from '@huayue/grpc-js'
import protoLoader from '@grpc/proto-loader'
import { AutoRouter, cors, error, json } from 'itty-router'
import dotenv from 'dotenv'
import path, { dirname } from 'path'
import { fileURLToPath } from 'url'
import { createServerAdapter } from '@whatwg-node/server'
import { createServer } from 'http'

// 加载环境变量
dotenv.config();

// 获取当前文件的目录路径（ESM 方式）
const __dirname = dirname(fileURLToPath(import.meta.url));

// 初始化配置
class Config {
  constructor() {
    this.API_PREFIX = process.env.API_PREFIX || '/'
    this.API_KEY = process.env.API_KEY || ''
    this.MAX_RETRY_COUNT = process.env.MAX_RETRY_COUNT || 3
    this.RETRY_DELAY = process.env.RETRY_DELAY || 5000
    this.COMMON_GRPC = 'runtime-native-io-vertex-inference-grpc-service-lmuw6mcn3q-ul.a.run.app'
    this.COMMON_PROTO = path.join(__dirname, '..', 'protos', 'VertexInferenceService.proto')
    this.GPT_GRPC = 'runtime-native-io-gpt-inference-grpc-service-lmuw6mcn3q-ul.a.run.app'
    this.GPT_PROTO = path.join(__dirname, '..', 'protos', 'GPTInferenceService.proto')
    this.PORT = process.env.PORT || 8787
    // 添加支持的模型列表
    this.SUPPORTED_MODELS = process.env.SUPPORTED_MODELS || [
      'gpt-4o-mini',
      'gpt-4o',
      'gpt-4-turbo',
      'gpt-4',
      'gpt-3.5-turbo',
      'claude-3-sonnet@20240229',
      'claude-3-opus@20240229',
      'claude-3-haiku@20240307',
      'claude-3-5-sonnet@20240620',
      'gemini-1.5-flash',
      'gemini-1.5-pro',
      'chat-bison',
      'codechat-bison',
    ]
  }

  // 添加模型验证方法
  isValidModel(model) {
    // 处理 Claude 模型的特殊格式
    const RegexInput = /^(claude-3-(5-sonnet|haiku|sonnet|opus))-(\d{8})$/
    const matchInput = model.match(RegexInput)
    const normalizedModel = matchInput ? `${matchInput[1]}@${matchInput[3]}` : model

    return this.SUPPORTED_MODELS.includes(normalizedModel)
  }
}

class GRPCHandler {
  constructor(protoFilePath) {
    // 动态加载传入的 .proto 文件路径
    this.packageDefinition = protoLoader.loadSync(protoFilePath, {
      keepCase: true,
      longs: String,
      enums: String,
      defaults: true,
      oneofs: true,
    })
  }
}
const config = new Config()

// 中间件
// 添加运行回源
const { preflight, corsify } = cors({
  origin: '*',
  allowMethods: '*',
  exposeHeaders: '*',
})

// 添加认证
const withAuth = (request) => {
  if (config.API_KEY) {
    const authHeader = request.headers.get('Authorization')
    if (!authHeader || !authHeader.startsWith('Bearer ')) {
      return error(401, 'Unauthorized: Missing or invalid Authorization header')
    }
    const token = authHeader.substring(7)
    if (token !== config.API_KEY) {
      return error(403, 'Forbidden: Invalid API key')
    }
  }
}

// 返回运行信息
const logger = (res, req) => {
  console.log(req.method, res.status, req.url, Date.now() - req.start, 'ms')
}

const router = AutoRouter({
  before: [preflight],
  missing: () => error(404, '404 not found.'),
  finally: [corsify, logger],
})

// Router路径
router.get('/', () => json({
    service: "AI Chat Completion Proxy",
    usage: {
        endpoint: "/v1/chat/completions",
        method: "POST",
        headers: {
            "Content-Type": "application/json",
            "Authorization": "Bearer YOUR_API_KEY"
        },
        body: {
            model: "One of: gpt-4o-mini, gpt-4o, gpt-4-turbo, gpt-4, gpt-3.5-turbo, claude-3-sonnet-20240229, claude-3-opus-20240229, claude-3-haiku-20240307, claude-3-5-sonnet-20240620, gemini-1.5-flash, gemini-1.5-pro, chat-bison, codechat-bison",
            messages: [
                { role: "system", content: "You are a helpful assistant." },
                { role: "user", content: "Hello, who are you?" }
            ],
            stream: false,
            temperature: 0.7,
            top_p: 1
        }
    },
    availableModels: [
        "gpt-4o-mini",
        "gpt-4o",
        "gpt-4-turbo",
        "gpt-4",
        "gpt-3.5-turbo",
        "claude-3-sonnet-20240229",
        "claude-3-opus-20240229",
        "claude-3-haiku-20240307",
        "claude-3-5-sonnet-20240620",
        "gemini-1.5-flash",
        "gemini-1.5-pro",
        "chat-bison",
        "codechat-bison"
    ],
    note: "Replace YOUR_API_KEY with your actual API key."
}));

router.get(config.API_PREFIX + '/v1/models', withAuth, () =>
  json({
    object: 'list',
    data: [
      { id: 'gpt-4o-mini', object: 'model', owned_by: 'pieces-os' },
      { id: 'gpt-4o', object: 'model', owned_by: 'pieces-os' },
      { id: 'gpt-4-turbo', object: 'model', owned_by: 'pieces-os' },
      { id: 'gpt-4', object: 'model', owned_by: 'pieces-os' },
      { id: 'gpt-3.5-turbo', object: 'model', owned_by: 'pieces-os' },
      { id: 'claude-3-sonnet-20240229', object: 'model', owned_by: 'pieces-os' },
      { id: 'claude-3-opus-20240229', object: 'model', owned_by: 'pieces-os' },
      { id: 'claude-3-haiku-20240307', object: 'model', owned_by: 'pieces-os' },
      { id: 'claude-3-5-sonnet-20240620', object: 'model', owned_by: 'pieces-os' },
      { id: 'gemini-1.5-flash', object: 'model', owned_by: 'pieces-os' },
      { id: 'gemini-1.5-pro', object: 'model', owned_by: 'pieces-os' },
      { id: 'chat-bison', object: 'model', owned_by: 'pieces-os' },
      { id: 'codechat-bison', object: 'model', owned_by: 'pieces-os' },
    ],
  }),
)

router.post(config.API_PREFIX + '/v1/chat/completions', withAuth, (req) => handleCompletion(req));

async function GrpcToPieces(inputModel, OriginModel, message, rules, stream, temperature, top_p) {
  // 在非GPT类型的模型中，temperature和top_p是无效的
  // 使用系统的根证书
  const credentials = grpc.credentials.createSsl()
  let client, request
  if (inputModel.includes('gpt')) {
    // 加载proto文件
    const packageDefinition = new GRPCHandler(config.GPT_PROTO).packageDefinition
    // 构建请求消息
    request = {
      models: inputModel,
      messages: [
        { role: 0, message: rules }, // system
        { role: 1, message: message }, // user
      ],
      temperature: temperature || 0.1,
      top_p: top_p ?? 1,
    }
    // 获取gRPC对象
    const GRPCobjects = grpc.loadPackageDefinition(packageDefinition).runtime.aot.machine_learning.parents.gpt
    client = new GRPCobjects.GPTInferenceService(config.GPT_GRPC, credentials)
  } else {
    // 加载proto文件
    const packageDefinition = new GRPCHandler(config.COMMON_PROTO).packageDefinition
    // 构建请求消息
    request = {
      models: inputModel,
      args: {
        messages: {
          unknown: 1,
          message: message,
        },
        rules: rules,
      },
    }
    // 获取gRPC对象
    const GRPCobjects = grpc.loadPackageDefinition(packageDefinition).runtime.aot.machine_learning.parents.vertex
    client = new GRPCobjects.VertexInferenceService(config.COMMON_GRPC, credentials)
  }
  return await ConvertOpenai(client, request, inputModel, OriginModel, stream)
}

async function messagesProcess(messages) {
  let rules = ''
  let message = ''

  for (const msg of messages) {
    let role = msg.role
    // 格式化为字符串
    const contentStr = Array.isArray(msg.content)
      ? msg.content
          .filter((item) => item.text)
          .map((item) => item.text)
          .join('') || ''
      : msg.content
    // 判断身份
    if (role === 'system') {
      rules += `system:${contentStr};\r\n`
    } else if (['user', 'assistant'].includes(role)) {
      message += `${role}:${contentStr};\r\n`
    }
  }

  return { rules, message }
}

async function ConvertOpenai(client, request, inputModel, OriginModel, stream) {
  const metadata = new grpc.Metadata()
  metadata.set('User-Agent', 'dart-grpc/2.0.0')
  for (let i = 0; i < config.MAX_RETRY_COUNT; i++) {
    try {
      if (stream) {
        const call = client.PredictWithStream(request, metadata)
        const encoder = new TextEncoder()
        const ReturnStream = new ReadableStream({
          start(controller) {
            // 处理数据
            call.on('data', (response) => {
              try {
                let response_code = Number(response.response_code)
                if (response_code === 204) {
                  controller.close()
                  call.destroy()
                } else if (response_code === 200) {
                  let response_message
                  if (inputModel.includes('gpt')) {
                    response_message = response.body.message_warpper.message.message
                  } else {
                    response_message = response.args.args.args.message
                  }
                  controller.enqueue(
                    encoder.encode(`data: ${JSON.stringify(ChatCompletionStreamWithModel(response_message, OriginModel))}\n\n`),
                  )
                } else {
                  console.error(`Invalid response code: ${response_code}`)
                  controller.error(error)
                }
              } catch (error) {
                console.error('Error processing stream data:', error)
                controller.error(error)
              }
            })

            // 处理错误
            call.on('error', (error) => {
              console.error('Stream error:', error)
              // 如果是 INTERNAL 错误且包含 RST_STREAM，可能是正常的流结束
              if (error.code === 13 && error.details.includes('RST_STREAM')) {
                controller.close()
              } else {
                controller.error(error)
              }
              call.destroy()
            })

            // 处理结束
            call.on('end', () => {
              controller.close()
            })

            // 处理取消
            return () => {
              call.destroy()
            }
          },
        })
        return new Response(ReturnStream, {
          headers: {
            'Content-Type': 'text/event-stream',
            Connection: 'keep-alive',
            'Cache-Control': 'no-cache',
            'Transfer-Encoding': 'chunked',
          },
        })
      } else {
        // 非流式调用保持不变
        const call = await new Promise((resolve, reject) => {
          client.Predict(request, metadata, (err, response) => {
            if (err) reject(err)
            else resolve(response)
          })
        })
        let response_code = Number(call.response_code)
        if (response_code === 200) {
          let response_message
          if (inputModel.includes('gpt')) {
            response_message = call.body.message_warpper.message.message
          } else {
            response_message = call.args.args.args.message
          }
          return new Response(JSON.stringify(ChatCompletionWithModel(response_message, OriginModel)), {
            headers: {
              'Content-Type': 'application/json',
            },
          })
        } else {
          throw new Error(`Invalid response code: ${response_code}`)
        }
      }
    } catch (err) {
      console.error(`Attempt ${i + 1} failed:`, err)
      await new Promise((resolve) => setTimeout(resolve, config.RETRY_DELAY))
    }
  }
  return new Response(
    JSON.stringify({
      error: {
        message: 'An error occurred while processing your request',
        type: 'server_error',
        code: 'internal_error',
        param: null,
      },
    }),
    {
      status: 500,
      headers: {
        'Content-Type': 'application/json',
      },
    },
  )
}

function ChatCompletionWithModel(message, model) {
  return {
    id: 'Chat-Nekohy',
    object: 'chat.completion',
    created: Date.now(),
    model,
    usage: {
      prompt_tokens: 0,
      completion_tokens: 0,
      total_tokens: 0,
    },
    choices: [
      {
        message: {
          content: message,
          role: 'assistant',
        },
        index: 0,
      },
    ],
  }
}

function ChatCompletionStreamWithModel(text, model) {
  return {
    id: 'chatcmpl-Nekohy',
    object: 'chat.completion.chunk',
    created: 0,
    model,
    choices: [
      {
        index: 0,
        delta: {
          content: text,
        },
        finish_reason: null,
      },
    ],
  }
}

async function handleCompletion(request) {
  try {
    // 解析openai格式API请求
    const { model: OriginModel, messages, stream, temperature, top_p } = await request.json()
    const RegexInput = /^(claude-3-(5-sonnet|haiku|sonnet|opus))-(\d{8})$/
    const matchInput = OriginModel.match(RegexInput)
    const inputModel = matchInput ? `${matchInput[1]}@${matchInput[3]}` : OriginModel
    // 添加模型验证
    if (!config.isValidModel(inputModel)) {
      return new Response(
        JSON.stringify({
          error: {
            message: `Model '${OriginModel}' does not exist`,
            type: 'invalid_request_error',
            param: 'model',
            code: 'model_not_found',
          },
        }),
        {
          status: 404,
          headers: {
            'Content-Type': 'application/json',
          },
        },
      )
    }
    console.log(inputModel, messages, stream)
    // 解析system和user/assistant消息
    const { rules, message: content } = await messagesProcess(messages)
    console.log(rules, content)
    // 响应码，回复的消息
    return await GrpcToPieces(inputModel, OriginModel, content, rules, stream, temperature, top_p)
  } catch (err) {
    return error(500, err.message)
  }
}

;(async () => {
  //For Cloudflare Workers
  if (typeof addEventListener === 'function') return
  // For Nodejs
  const ittyServer = createServerAdapter(router.fetch)
  console.log(`Listening on http://localhost:${config.PORT}`)
  const httpServer = createServer(ittyServer)
  httpServer.listen(config.PORT)
})()