HuggingFace

To use HuggingFace embeddings locally, you can use the @xenova/transformers package with Settings.embedFunc.

Installation

npm

npm i @xenova/transformers

pnpm

pnpm add @xenova/transformers

yarn

yarn add @xenova/transformers

bun

bun add @xenova/transformers

Basic Setup

import { Settings, TextEmbedFunc } from "@vectorstores/core";
import { pipeline } from "@xenova/transformers";

// Create embedding function using HuggingFace transformers
async function createHuggingFaceEmbedding(
  modelType: string = "Xenova/all-MiniLM-L6-v2",
): Promise<TextEmbedFunc> {
  const extractor = await pipeline("feature-extraction", modelType);

  return async (input) => {
    const texts = Array.isArray(input) ? input : [input];
    const embeddings: number[][] = [];

    for (const text of texts) {
      const output = await extractor(text, { pooling: "mean", normalize: true });
      embeddings.push(Array.from(output.data));
    }

    return embeddings;
  };
}

// Usage
Settings.embedFunc = await createHuggingFaceEmbedding();

Available Models

By default, Xenova/all-MiniLM-L6-v2 is used. Other popular models include:

• Xenova/all-MiniLM-L6-v2 - Fast and lightweight, good general purpose
• Xenova/bge-small-en-v1.5 - Higher quality for English text
• Xenova/all-mpnet-base-v2 - Better quality, slightly slower

These models run entirely locally in your JavaScript runtime, making them ideal for privacy-sensitive applications or offline use.