dexaai / openai-fetch Goto Github PK

Thanks for this API wrapper !

Calls to the model endpoint (i.e. call GET https://api.openai.com/v1/models) is not included in this wrapper.

It's not critical, but could be nice (type safety and error management).

Openai-node no longer uses axios; it uses node-fetch by default in Node, and builtin fetch in all other envs.

Wanted to run API calls from edge, and had to find fetch-based client.

Stumbled upon this repo, but when searching npmjs.org, I found 2 packages pointing to this repo: openai-fetch and @bytemain/openai-fetch. Found this a bit strange/questioning.

Is this function already implemented?

https://platform.openai.com/docs/guides/moderation

Nice!

Can this be relied upon for production, or might anything break it at the openai end?

Thanks

Opening an issue to track this feature, since it seems to keep coming up, and I'm expecting to make use of it soon for a project.

See also:

• https://www.beskar.co/blog/streaming-openai-completions-vercel-edge
• https://www.npmjs.com/package/@beskar-labs/parse-json-sse
• https://github.com/transitive-bullshit/chatgpt-api/blob/main/src/fetch-sse.ts

Seeing the following error when trying to load an Open AI completion from a Next.js 13 page (server component).

import createOpenAiClient from '~/lib/openai';

async function getData() {
  const data = await createOpenAiClient().createCompletion({
    prompt: `List some cool objects in a room`,
    model: 'text-curie-001',
    temperature: 0,
    frequency_penalty: 0.1,
    max_tokens: 256,
  });

  console.log(JSON.stringify(data.response, null, 2));

  return data.completion;
}

export default async function Page() {
  const response = await getData();

To create the client, I originally used the code cited in the README. However, after seeing this error, I tried to create the client on-demand:

import { OpenAIClient } from 'openai-fetch';

function createOpenAiClient() {
  return new OpenAIClient({ apiKey: process.env.OPENAI_API_KEY });
}

export default createOpenAiClient;

But to no avail.

I noticed you're using ky. A cursory search yielded a seemingly related issue: sindresorhus/ky#479

Currently, the library doesn't support any interaction with fine tunings.

The following is a list of wanted methods to be included:

• retrieveFineTune
• listFineTunes
• createFineTune
• cancelFineTune
• listFineTuneEvents

Thanks for your efforts.

it is a ky's issue.

• sindresorhus/ky#366

it seem that CloudFlare Workers hasn't implements the property.

The README on this project says this:

Unfortunately, the official openai-node uses Axios, which only supports Node and is bloated.

Reasons to consider using openai-fetch:
Supports all envs with native fetch: Node 18+, browsers, Deno, Cloudflare Workers, etc

As of https://github.com/openai/openai-node/releases/tag/v4.0.0, openai uses the native fetch by default in Deno, Cloudflare workers, etc. It does use node-fetch in Node 18+, though may be moving to undici there soon.

The package download size stats may have changed as well, though most of the size of the official package is types, sourcemaps, etc which aren't imported at runtime. The runtime impact is usually around 25kb (still larger than this package).

Please can you elaborate more on this phrase in readme?

Unfortunately, the official [openai-node](https://github.com/openai/openai-node) uses Axios, which only supports Node and is bloated.

I got curious why you say that the axios and nodejs client is "bloated"

Hi, first of all, thanks for great lib

These lines would cause an error in environments where process is not available (ex. Cloudflare or browsers) in case if apiKey or organizationId is not specified

Error I get when trying to run in Cloudflare worker

✘ [ERROR] Uncaught (in promise) ReferenceError: process is not defined

  const organizationId = opts.organizationId || process.env.OPENAI_ORG_ID;
                                                ^
      at OpenAIClient
  (/root/dev/app/node_modules/openai-fetch/src/openai-client.ts:62:50)
      at fetch (/root/dev/app/src/worker.ts:32:17)

As lib claims it supports all environments, including browsers, process should not be used as fallback

Is it possible to extend API configuration to pass a specific fetch library?

It might solve some issues related to the different hosting environments.

something like that:

const client = new OpenAIClient({ apiKey: '<your api key>', fetch: myFetch });