ChatOpenAI breaks edge streaming on Vercel Edge functions.

[oalexdoda]

ChatOpenAI breaks edge streaming on Vercel Edge functions.

Error:

../../node_modules/.pnpm/openai@4.4.0/node_modules/openai/core.mjs
A Node.js API is used (process.arch at line: 531) which is not supported in the Edge Runtime.

Import trace for requested module:

../../node_modules/.pnpm/openai@4.4.0/node_modules/openai/core.mjs
../../node_modules/.pnpm/openai@4.4.0/node_modules/openai/index.mjs
../../node_modules/.pnpm/langchain@0.0.175_axios@1.6.0_mammoth@1.6.0/node_modules/langchain/dist/chat_models/openai.js
../../node_modules/.pnpm/langchain@0.0.175_axios@1.6.0_mammoth@1.6.0/node_modules/langchain/chat_models/openai.js

...

The import causing the error:
import { ChatOpenAI } from 'langchain/chat_models/openai';

Potential Solution:
If the native OpenAI library doesn't support edge streaming / functions, then LangChain should implement that chat module via a native fetch.

[jacoblee93]

What bundler are you using?

Have noticed this warning in our tests but we use this live in production with Webpack + Next pretty frequently, so I assumed the bundler was smart enough to filter it out:

https://github.com/langchain-ai/langchain-nextjs-template

[oalexdoda]

What bundler are you using?

Have noticed this warning in our tests but we use this live in production with Webpack + Next pretty frequently, so I assumed the bundler was smart enough to filter it out:

https://github.com/langchain-ai/langchain-nextjs-template

Next.js with Webpack, but I have this custom config - not sure if it matters:

const { withDev } = require('@destacked/dev');
const { withSentryConfig } = require('@sentry/nextjs');

const withBundleAnalyzer = require('@next/bundle-analyzer')({
	enabled: process.env.ANALYZE === 'true',
	openAnalyzer: false
});

const config = {
	webpack: (config) => {
		if (process.env.NODE_ENV === 'development') {
			config.watchOptions = {
				poll: 10000
			};
		}

		config.module.rules.push({
			test: /\.XXXXXXXXXX.txt$/,
			type: 'asset/source'
		});

		return config;
	},
	experimental: {
		useDeploymentId: true,
		webpackBuildWorker: true
	},
	transpilePackages: [
		'@destacked/js',
		'@destacked/server',
		'@destacked/interface',
		'@destacked/layouts',
		'@destacked/apps-gleap',
		'@destacked/apps-mirror',
		'@tabler/icons-react'
	],
	poweredByHeader: false,
	sentry: {
		hideSourceMaps: true,
		widenClientFileUpload: true
	},
	images: {
		dangerouslyAllowSVG: true,
		remotePatterns: [
			{ hostname: 'XXXXXXXXXX' }
		]
	},
	async rewrites() {
		return [
			{
				source: '/XXXXXXXXXX',
				destination: 'XXXXXXXXXX'
			}
		];
	},
	async redirects() {
		return [
			{
				source: '/XXXXXXXXXX',
				destination: 'XXXXXXXXXX',
				permanent: false
			},
		];
	}
};

module.exports = withSentryConfig(withBundleAnalyzer(withDev(config)), {
	org: 'XXXXXXXXXX',
	project: 'XXXXXXXXXX',
	silent: true
});

[jacoblee93]

CC @rattrayalex is it possible to remove references like this?

[rattrayalex]

Hey @jacoblee93 , thanks for the ping. We don't actually call that at runtime on that platform, so I'd consider it a bug/problem with the import tracer… but I can see if I can move that code into another file to make this easier.

[rattrayalex]

Ah, I see the user is on openai@4.4.0, can you try with a later version? We've made improvements to related areas and I wonder if it'd be fixed.

Also, I'm a bit surprised to see .cjs used in a Next build instead of .mjs, know what might be up with that?

[jacoblee93]

Yes, definitely a bundler/chunking issue.

Ok will try bumping the version! Thanks for the look.

[jacoblee93]

Please reopen if this persists past the next release (0.0.180)