@ezijing/ai-core

@ezijing/ai-core 提供按模型路由的多 provider client。

安装

npm install @ezijing/ai-core

创建 Provider

import {
  createDeepSeekProvider,
  createQianfanProvider,
  createVolcengineProvider,
} from '@ezijing/ai-core'

const deepseek = createDeepSeekProvider({ apiKey: 'your-deepseek-key' })
const volcengine = createVolcengineProvider({ apiKey: 'your-volcengine-key' })
const qianfan = createQianfanProvider({ apiKey: 'your-qianfan-key' })

创建 Client

import { createClient } from '@ezijing/ai-core'

const client = createClient({
  providers: {
    deepseek,
    volcengine,
    qianfan,
  },
})

获取模型列表

const models = client.getModels()

模型结构：

type AIModel = {
  label: string
  provider: string
  model: string
  value: string
  capabilities: ('text' | 'image' | 'video')[]
  disabled?: boolean
}

其中 disabled: true 的模型不会出现在 client.getModels() 里。

示例：

[
  {
    label: 'DeepSeek-V3',
    provider: 'deepseek',
    model: 'deepseek-chat',
    value: 'deepseek:deepseek-chat',
    capabilities: ['text'],
  },
]

文本生成

const response = await client.generateText({
  model: 'deepseek:deepseek-chat',
  prompt: '你好，请介绍一下自己',
})

console.log(response.content)
console.log(response.reasoning)
console.log(response.references)

流式文本

const result = client.streamText({
  model: 'volcengine:deepseek-v3-2-251201',
  prompt: '请写一个故事',
})

for await (const chunk of result.textStream) {
  console.log(chunk)
}

const finalText = await result.text
const response = await result.response

消费完整流：

for await (const chunk of result.fullStream) {
  console.log(chunk.content)
  console.log(chunk.contentDelta)
  console.log(chunk.reasoning)
}

图片生成

const image = await client.generateImage({
  model: 'volcengine:doubao-seedream-4-5-251128',
  prompt: '一只可爱的小猫',
})

console.log(image.images)

视频生成

const result = client.streamVideo({
  model: 'volcengine:doubao-seedance-1-5-pro-251215',
  prompt: '一只可爱的小猫在花园里玩耍',
})

for await (const update of result.statusStream) {
  console.log(update.status)
}

const video = await result.response
console.log(video.url)
console.log(video.duration)
console.log(video.ratio)
console.log(video.resolution)

取消视频任务：

result.cancel()

RequestOptions

所有方法都支持统一的第二个参数：

type RequestOptions = AxiosRequestConfig

最常用的是 signal：

const controller = new AbortController()

const result = client.streamText(
  {
    model: 'deepseek:deepseek-chat',
    prompt: '你好',
  },
  {
    signal: controller.signal,
  },
)

controller.abort()

路由规则

使用 provider:model 作为唯一模型值
client 只负责注册 provider 和按模型路由
provider 自己负责创建和请求适配

注意事项

浏览器直连时，provider key 仍然会暴露给最终用户