npm package discovery and stats viewer.

Discover Tips

  • General search

    [free text search, go nuts!]

  • Package details

    pkg:[package-name]

  • User packages

    @[username]

Sponsor

Optimize Toolset

I’ve always been into building performant and accessible sites, but lately I’ve been taking it extremely seriously. So much so that I’ve been building a tool to help me optimize and monitor the sites that I build to make sure that I’m making an attempt to offer the best experience to those who visit them. If you’re into performant, accessible and SEO friendly sites, you might like it too! You can check it out at Optimize Toolset.

About

Hi, 👋, I’m Ryan Hefner  and I built this site for me, and you! The goal of this site was to provide an easy way for me to check the stats on my npm packages, both for prioritizing issues and updates, and to give me a little kick in the pants to keep up on stuff.

As I was building it, I realized that I was actually using the tool to build the tool, and figured I might as well put this out there and hopefully others will find it to be a fast and useful way to search and browse npm packages as I have.

If you’re interested in other things I’m working on, follow me on Twitter or check out the open source projects I’ve been publishing on GitHub.

I am also working on a Twitter bot for this site to tweet the most popular, newest, random packages from npm. Please follow that account now and it will start sending out packages soon–ish.

Open Software & Tools

This site wouldn’t be possible without the immense generosity and tireless efforts from the people who make contributions to the world and share their work via open source initiatives. Thank you 🙏

© 2024 – Pkg Stats / Ryan Hefner

not-a-vector-database

v0.0.1

Published

Store and search embeddings

Downloads

10

Readme

You need a few lines of JS, not a vector database

Suppose you are implementing RAG for your AI app, and you have used web APIs or an inference engine to generate a large number of embeddings, and now you need to find out the best results matching a query embedding, what do you do?

Use a vector database? No, you only need a few lines of JavaScript.

A quick introduction to node-mlx

MLX is a full-featured machine learning framework, with easy-to-understand source code and small binary sizes. And node-mlx is the JavaScript binding of it.

MLX only has GPU support for macOS, and but its CPU support, implemented with vectorized instructions, is still fast on Linux.

import {core as mx, nn} from '@frost-beta/mlx';

Embeddings search, in a few lines of JS

Suppose you want to find out the results with highest similaries to query, from the embeddings.

const embeddings = [
    [  0.037035,  0.0760545, ... ],
    [  0.034029, -0.0227216, ... ],
    ...
    [ -0.028612,  0.0052857, ... ],
];
const query = [ -0.019773, 0.006021, ... ];

With node-mlx, you can use the builtin nn.losses.cosineSimilarityLoss API to do the search.

const embeddingsTensor = mx.array(embeddings);
const queryTensor = mx.array([ query ]);
const scoresTensor = nn.losses.cosineSimilarityLoss(queryTensor, embeddingsTensor);
const scores: Float32Array = scoresTensor.toTypedArray();

The scores array stores the cosine similarities between the query and embeddings.

(If you are wondering how we can compute cosine similarities between a 1x1 tensor and a 1xN tensor, it is called broadcasting.)

console.log(scores);
// [ 0.233244, 0.012492, ..., 0.43298 ]

Sorting

Once you get the scores array, you can use the usual JavScript code to filter and sort the results. But you can also use MLX if the number of results is large enough to make JavScript engine struggle.

// Get the top 10 scores.
const topTen = mx.topk(scoresTensor, 10);
console.log(topTen.toTypedArray());
// [ 0.894323, 0.594729, ... ]

// Sort the scores.
let sortedScores = mx.sort(scoresTensor);
console.log(sortedScores.toTypedArray());
// [ 0.01287, 0.1502876, ... ]
sortedScores = sortedScores.index(mx.Slice(null, null, -1));
console.log(sortedScores.toTypedArray());
// [ 0.894323, 0.594729, ... ]

// Get the indices of the scores ordered by their values in the array.
const indices = mx.argsort(scoresTensor)
    .index(mx.Slice(null, null, -1))
    .toTypedArray();
console.log(indices);
// [ 8, 9, ... ]
console.log(indices.map(i => scores[i]));
// [ 0.894323, 0.594729, ... ]

The array.index(mx.Slice(null, null, -1)) code looks alien, it is actually the JavaScript version of Python's array[::-1], which reverse the array. You can of course convert the result to JavaScript Array frist and then call reverse(), but it would be slower if the array is very large.

A Node.js module

If after reading above introductions you still find MXL cumbersome to use (which is normal if you had zero experience with NumPy or PyTorch), I have wrapped the code into a very simple Node.js module, which you can use to replace vector databases in many cases.

Install:

npm install not-a-vector-database

APIs:

export type EmbeddingInput = number[] | TypedArray;

export interface SearchOptions {
    /**
     * Results with scores larger than this value will be returned. Should be
     * between -1 and 1. Default is 0.
     */
    minimumScore?: number;
    /**
     * Restrict the number of results to return, default is 16.
     */
    maximumResults?: number;
}

export interface SearchResult {
    score: number;
    data: unknown;
}

/**
 * In-memory storage of embeddings and associated data.
 */
export class Storage {
    embeddings?: mx.array;
    data: unknown[];

    /**
     * Initialize from the buffer.
     */
    loadFromBuffer(buffer: Buffer): void;

    /**
     * Dump the data to a buffer.
     */
    dumpToBuffer(): Buffer;

    /**
     * Add data to the storage.
     */
    push(...items: {embedding: EmbeddingInput; data: unknown;}[]): void;

    /**
     * Return the data which are most relevant to the embedding.
     */
    search(embedding: EmbeddingInput, options?: SearchOptions): SearchResult[];
}

Example:

import {Storage} from 'not-a-vector-database';

const storage = new Stroage();
storage.push({embedding, data: 'some data'});

const results = storage.search(embedding);

fs.writeFileSync('storage.bser', storage.dumpToBuffer());
storage.loadFromBuffer(fs.readFileSync('storage.bser'));

There is also a benchmark.ts script that you can use to test the performance. On my 2018 Intel MacBook Pro which has no GPU support, searching from 1 million embeddings with size of 128 takes about 900ms.