Nuxt - Streaming Issue with the example "Nuxt OpenAI Starter" on Vercel

[JuBertoo]

I've deployed the example "Nuxt OpenAI Starter" on Vercel and I'm encountering a streaming problem. Instead of sending data progressively, the system waits for the server's response to be complete before displaying any data. This issue doesn't occur when I run it locally.

https://nuxt-openai-vert.vercel.app/

Could this be related to Vercel's configuration? Any help is appreciated.

[jaredpalmer]

Nuxt doesn't support streaming at the moment. You can track the issue here: https://github.com/unjs/nitro/issues/1327

[Hebilicious]

@jaredpalmer This should currently be working with node runtimes, but support for (edge) streaming landed with h3 1.7.0, here's a working example with cloudflare pages :

https://github.com/Hebilicious/vercel-sdk-ai/blob/cloudflare-official/examples/nuxt-openai/server/api/chat.ts

(I haven't tried it on vercel-edge but this should work too)

Note that this custom sendStream utility will be provided by the framework soon.

@JuBertoo install h3 1.7.0, and update your code to do something like this :

export default defineEventHandler(async (event: any) => {
  // Extract the `prompt` from the body of the request
  const { messages } = await readBody(event)

  // Ask OpenAI for a streaming chat completion given the prompt
  const response = await openai.createChatCompletion({
    model: 'gpt-3.5-turbo',
    stream: true,
    messages: messages.map((message: any) => ({
      content: message.content,
      role: message.role
    }))
  })

  // Convert the response into a friendly text-stream
  const stream = OpenAIStream(response)
  // Respond with the stream
  return sendStream(event, stream)
})

function sendStream(event: H3Event, stream: ReadableStream) {
  // Mark to prevent h3 handling response
  event._handled = true

  // Workers (unenv)
  // @ts-expect-error _data will be there.
  event.node.res._data = stream

  // Node.js
  if (event.node.res.socket) {
    stream.pipeTo(
      new WritableStream({
        write(chunk) {
          event.node.res.write(chunk)
        },
        close() {
          event.node.res.end()
        }
      })
    )
  }
}

[JuBertoo]

Thank you for your response, @Hebilicious , but it still doesn't work... I have installed the dependency 'h3': '^1.7.0' and updated the code. It may not be compatible with Vercel's Edge Functions.

import { Configuration, OpenAIApi } from 'openai-edge';
import { OpenAIStream } from 'ai';
import type { H3Event } from 'h3';

// Create an OpenAI API client (that's edge friendly!)
const config = new Configuration({
  // eslint-disable-next-line react-hooks/rules-of-hooks
  apiKey: useRuntimeConfig().openaiApiKey,
});
const openai = new OpenAIApi(config);

export default defineEventHandler(async (event: any) => {
  // Extract the `prompt` from the body of the request
  const { messages } = await readBody(event);

  // Ask OpenAI for a streaming chat completion given the prompt
  const response = await openai.createChatCompletion({
    model: 'gpt-3.5-turbo',
    stream: true,
    messages: messages.map((message: any) => ({
      content: message.content,
      role: message.role,
    })),
  });

  // Convert the response into a friendly text-stream
  const stream = OpenAIStream(response);
  // Respond with the stream
  return sendStream(event, stream);

  function sendStream(event: H3Event, stream: ReadableStream) {
    // Mark to prevent h3 handling response
    event._handled = true;

    // Workers (unenv)
    event.node.res._data = stream;

    // Node.js
    if (event.node.res.socket) {
      stream.pipeTo(
        new WritableStream({
          write(chunk) {
            event.node.res.write(chunk);
          },
          close() {
            event.node.res.end();
          },
        })
      );
    }
  }
});

[Hebilicious]

Thank you for your response, @Hebilicious , but it still doesn't work... I have installed the dependency 'h3': '^1.7.0' and updated the code. It may not be compatible with Vercel's Edge Functions.

import { Configuration, OpenAIApi } from 'openai-edge';
import { OpenAIStream } from 'ai';
import type { H3Event } from 'h3';

// Create an OpenAI API client (that's edge friendly!)
const config = new Configuration({
  // eslint-disable-next-line react-hooks/rules-of-hooks
  apiKey: useRuntimeConfig().openaiApiKey,
});
const openai = new OpenAIApi(config);

export default defineEventHandler(async (event: any) => {
  // Extract the `prompt` from the body of the request
  const { messages } = await readBody(event);

  // Ask OpenAI for a streaming chat completion given the prompt
  const response = await openai.createChatCompletion({
    model: 'gpt-3.5-turbo',
    stream: true,
    messages: messages.map((message: any) => ({
      content: message.content,
      role: message.role,
    })),
  });

  // Convert the response into a friendly text-stream
  const stream = OpenAIStream(response);
  // Respond with the stream
  return sendStream(event, stream);

  function sendStream(event: H3Event, stream: ReadableStream) {
    // Mark to prevent h3 handling response
    event._handled = true;

    // Workers (unenv)
    event.node.res._data = stream;

    // Node.js
    if (event.node.res.socket) {
      stream.pipeTo(
        new WritableStream({
          write(chunk) {
            event.node.res.write(chunk);
          },
          close() {
            event.node.res.end();
          },
        })
      );
    }
  }
});

I assume that since it runs on cloudflare it should run on vercel edge, I will try to deploy an example.

[dosstx]

@Hebilicious Any update on whether you got it working on Vercel edge?

[Hebilicious]

@Hebilicious Any update on whether you got it working on Vercel edge?

I've been able to deploy with the CLI (ie running vercel deploy) without any issues.

https://nuxt-openai-vercel-hebilicious.vercel.app/

Nuxt config

export default defineNuxtConfig({
  devtools: { enabled: true },
  modules: ['@nuxtjs/tailwindcss'],
  nitro: {
    preset: 'vercel-edge'
  },
  // You might not need it if you're not using pnpm
  alias: {
    'node:util': path.resolve(
      __dirname,
      'node_modules/unenv/runtime/node/util/index.cjs'
    ),
    'node:net': path.resolve(
      __dirname,
      'node_modules/unenv/runtime/node/net/index.cjs'
    )
  },
  runtimeConfig: {
    openaiApiKey: ''
  }
})

Server API

// ./api/chat.ts
import { Configuration, OpenAIApi } from 'openai-edge'
import { OpenAIStream } from 'ai'
import type { H3Event } from 'h3'

let openai: OpenAIApi

export default defineEventHandler(async (event: any) => {
// You can probably move this out of the event handler with vercel-edge
  if (!openai) {
    let apiKey = useRuntimeConfig().openaiApiKey as string
    const config = new Configuration({ apiKey })
    openai = new OpenAIApi(config)
  }

  // Extract the `prompt` from the body of the request
  const { messages } = await readBody(event)

  // Ask OpenAI for a streaming chat completion given the prompt
  const response = await openai.createChatCompletion({
    model: 'gpt-3.5-turbo',
    stream: true,
    messages: messages.map((message: any) => ({
      content: message.content,
      role: message.role
    }))
  })

  // Convert the response into a friendly text-stream
  const stream = OpenAIStream(response)
  // Respond with the stream
  return sendStream(event, stream)
})

// This will be provided by the framework in a future version
function sendStream(event: H3Event, stream: ReadableStream) {
  // Mark to prevent h3 handling response
  event._handled = true

  // Workers (unenv)
  // @ts-expect-error _data will be there.
  event.node.res._data = stream

  // Node.js
  if (event.node.res.socket) {
    stream.pipeTo(
      new WritableStream({
        write(chunk) {
          event.node.res.write(chunk)
        },
        close() {
          event.node.res.end()
        }
      })
    )
  }
}

Edit: Going through the UI, it looks like it's using edge functions properly.

@jaredpalmer What can we do from the Nuxt side to resolve this ? Update the example and add some information in the README for edge-functions caveats ?

[MaxLeiter]

@Hebilicious those seem like two good suggestions. Would you mind contributing a pull request?

[Giancarlo-Ma]

how can i intercept the streaming content to save it to db on cf worker

[lgrammel]

how can i intercept the streaming content to save it to db on cf worker

@Giancarlo-Ma you can add callback handlers to the OpenAIStream. Check out this example: https://sdk.vercel.ai/docs/guides/providers/openai#guide-save-to-database-after-completion