OpenAI Streams

Github | NPM | Docs

Now with ChatGPT API support! See Use with ChatGPT API. (Whisper coming soon!)

This library returns OpenAI API responses as streams only. Non-stream endpoints like edits etc. are simply a stream with only one chunk update.

• Prioritizes streams, so you can display a completion as it arrives.
• Auto-loads OPENAI_API_KEY from process.env.
• One single function with inferred parameter type based on the endpoint you provide.

Uses ReadableStream by default for browser, Edge Runtime, and Node 18+, with a NodeJS.Readable version available at openai-streams/node.

Installation

yarn add openai-streams
# -or-
npm i --save openai-streams

Usage

1. Set the OPENAI_API_KEY env variable (or pass the { apiKey } option).

   The library will throw if it cannot find an API key. Your program will load this at runtime from process.env.OPENAI_API_KEY by default, but you may override this with the { apiKey } option.

   IMPORTANT: For security, you should only load this from a process.env variable.

   await OpenAI(
     "completions", 
     {/* params */}, 
     { apiKey: process.env.MY_SECRET_API_KEY }
   )

2. Call the API via await OpenAI(endpoint, params).

   The params type will be inferred based on the endpoint you provide, i.e. for the "edits" endpoint, import('openai').CreateEditRequest will be enforced.

Edge/Browser: Consuming streams in Next.js Edge functions

This will also work in the browser, but you'll need users to paste their OpenAI key and pass it in via the { apiKey } option.

import { OpenAI } from "openai-streams";

export default async function handler() {
  const stream = await OpenAI(
    "completions",
    {
      model: "text-davinci-003",
      prompt: "Write a happy sentence.\n\n",
      max_tokens: 100
    },
  );

  return new Response(stream);
}

export const config = {
  runtime: "edge"
};

Node: Consuming streams in Next.js API Route (Node)

If you cannot use an Edge runtime or want to consume Node.js streams for another reason, use openai-streams/node:

import type { NextApiRequest, NextApiResponse } from "next";
import { OpenAI } from "openai-streams/node";

export default async function test (_: NextApiRequest, res: NextApiResponse) {
  const stream = await OpenAI(
    "completions",
    {
      model: "text-davinci-003",
      prompt: "Write a happy sentence.\n\n",
      max_tokens: 25
    }
  );

  stream.pipe(res);
}

See the example in example/src/pages/api/hello.ts. See also src/pages/api/demo.ts in nextjs-openai.

Use with ChatGPT API

By default, with mode = "tokens", you will receive just the message deltas. For full events, use mode = "raw".

See: https://platform.openai.com/docs/guides/chat/introduction

const stream = await OpenAI(
  "chat",
  {
    model: "gpt-3.5-turbo",
    messages: [
      { "role": "system", "content": "You are a helpful assistant that translates English to French." },
      { "role": "user", "content": "Translate the following English text to French: \"Hello world!\"" }
    ],
  },
);

In both modes, for Chat, you will receive a stream of serialized JSON objects. Even in mode = "tokens", you will need to parse the deltas because they sometimes indicate a role and sometimes indicate part of the message body. The stream chunks look like:

{"role":"assistant"}
{"content":"\""}
{"content":"Bonjour"}
{"content":" le"}
{"content":" monde"}
{"content":" !\""}
{}

Notes

1. Internally, streams are often manipulated using generators via for await (const chunk of yieldStream(stream)) { ... }. We recommend following this pattern if you find it intuitive.