Embedding values are different for the same model with the same parameters but different environments

[NikhilVerma]

System Info

"@huggingface/transformers": "3.0.2"
Bun v1.1.30-canary.28+2f8c20ef8 (macOS arm64)
Chrome Version 131.0.6778.70 arm64

Environment/Platform

• Website/web-app
• Browser extension
• Server-side (e.g., Node.js, Deno, Bun)
• Desktop app (e.g., Electron)
• Other (e.g., VSCode extension)

Description

Take this piece of code

import { FeatureExtractionPipeline, pipeline } from "@huggingface/transformers";

let embeddingPipeline: Promise<FeatureExtractionPipeline> | undefined;

export async function getEmbedding(str: string) {
	if (!embeddingPipeline) {
		// eslint-disable-next-line no-console
		console.log(`🧠 Loading embedding pipeline`);

		embeddingPipeline = pipeline("feature-extraction", "mixedbread-ai/mxbai-embed-large-v1", {
			device: "auto",
			dtype: "q8"
		});
	}

	const resolvedPipeline = await embeddingPipeline;

	const output = await resolvedPipeline(str, {
		pooling: "cls",
		normalize: true
	});

	return output.tolist()[0] as number[];
}

When I run this code in the Browser and try to embed the same string "Hello world!" vs when I embed it in the browser (using WASM) I get entirely different values

Using bun

[
  -0.0021202601492404938, 0.0469636432826519, 0.020890265703201294, -0.022750232368707657,
  -0.037959884852170944, 0.003005226841196418, 0.053498536348342896, 0.022222621366381645,
  0.04570423811674118, 0.01894044317305088, 0.013796147890388966, 0.020720647647976875,
  0.009724173694849014, 0.013014012947678566, -0.05031043663620949, -0.000107266751001589,
  -0.014651484787464142, -0.024391934275627136, -0.051498446613550186, 0.012894446961581707,

Using browser

 [
    -0.016504084691405296,
    -0.04971511662006378,
    -0.0009000280988402665,
    -0.02497180737555027,
    -0.024496370926499367,
    -0.007153023034334183,
    -0.004980664700269699,
    0.021212700754404068,
    0.027969202026724815,
    0.004761920776218176,
    0.02262801118195057,
    -0.012614945881068707,
    0.02732115238904953,
    -0.04224494472146034,
    -0.004798870999366045,
    -0.028831878677010536,
    -0.04931287840008736,
    -0.034003954380750656,
    -0.07406899333000183,
    0.003017101436853409,
    0.005531534552574158,
    -0.005652727093547583,
    -0.05290239304304123,
    -0.043359968811273575,

Reproduction

1. Embed a string using mixedbread-ai/mxbai-embed-large-v1 in console using Bun (Node should behave the same) (dtype resolves to cpu)
2. Embed the same string in browser (dtype resolves to wasm)
3. The values are different for the same piece of code

This makes any kind of RAG development impossible.

Update

1. Switching between wasm and webgpu on the browser also gives entirely different results (different from cpu). It seems like the behaviour is quite different with different backends. I don't think this should be the case

[NikhilVerma]

I have also tried ensuring that we use similar versions for both onnxruntime-web and node but the end-result is the same.

[NikhilVerma]

OK i found the fix. I think we should document it. Basically forcing CPU (instead of WASM) gives the desired consistency.

                 embeddingPipeline = pipeline("feature-extraction", LOCAL_EMBEDDING_MODEL_NAME, {
			device: "auto",
			dtype: "q8",
			session_options: {
				executionProviders: ["cpu"]
			}
		});

It's not ideal ofcoure, but since I run this in a workerpool I don't notice too much degradation