voy-search

Voy

A WASM vector similarity search engine written in Rust

• Tiny: 75KB gzipped, 69KB brotli.
• Fast: Create the best search experience for the users. Voy uses k-d tree to index and provide fast search
• Tree Shakable: Optimize bundle size and enable asynchronous capabilities for modern Web API, such as Web Workers.
• Resumable: Generate portable embeddings index anywhere, anytime.
• Worldwide: Designed to deploy and run on CDN edge servers.

🚜 Work in Progress

Voy is under active development. As a result, the API is not stable. Please be aware that there might be breaking changes before the upcoming 1.0 release.

A sneak peek of what we are working on:

• [ ] Built-in text transformation in WebAssembly: As of now, voy relies on JavaScript libraries like transformers.js to generate text embeddings. See Usage for more detail.
• [x] Index update: Currently it's required to re-build the index when a resource update occurs.
• [x] TypeScript support: Due to the limitation of WASM tooling, complex data types are not auto-generated.

Installation

# with npm
npm i voy-search

# with Yarn
yarn add voy-search

# with pnpm
pnpm add voy-search

APIs

class Voy

The Voy class encapsulates an index and exposes all the public methods Voy has to offer.

class Voy {
  /**
   * By instantiating with a resource, Voy will construct the index. If the resource is
   * absent, it will construct an empty index. Calling Voy.index() later on will override
   * the empty index.
   *
   * @param {Resource | undefined} resource
   */
  constructor(resource?: Resource);
  /**
   * Index given resource. Voy.index() is designed for the use case where a Voy instance
   * is instantiated without a resource. It will override the existing index. If you'd like
   * to keep the existing index, you can use Voy.add() to add your resource to the index.
   *
   * @param {Resource} resource
   */
  index(resource: Resource): void;
  /**
   * Search top k results with given query embedding.
   *
   * @param {Float32Array} query: Query Embedding
   * @param {number} k: Number of items in the search result
   * @returns {SearchResult}
   */
  search(query: Float32Array, k: number): SearchResult;
  /**
   * Add given resource to the index.
   *
   * @param {Resource} resource
   */
  add(resource: Resource): void;
  /**
   * Remove given resource from the index.
   *
   * @param {Resource} resource
   */
  remove(resource: Resource): void;
  /**
   * Remove all resources from the index.
   */
  clear(): void;
}

interface Resource {
  embeddings: Array<{
    id: string; // id of the resource
    title: string; // title of the resource
    url: string; // url to the resource
    embeddings: number[]; // embeddings of the resource
  }>;
}

interface SearchResult {
  neighbors: Array<{
    id: string; // id of the resource
    title: string; // title of the resource
    url: string; // url to the resource
  }>;
}

Individual Functions

Besides the Voy class, Voy also exports all the instance methods as individual functions.

index(resource: Resource): SerializedIndex

It indexes the given resource and returns a serialized index.

Parameters

interface Resource {
  embeddings: Array<{
    id: string; // id of the resource
    title: string; // title of the resource
    url: string; // url to the resource
    embeddings: number[]; // embeddings of the resource
  }>;
}

Return

type SerializedIndex = string;

search(index: SerializedIndex, query: Query, k: NumberOfResult): SearchResult

It deserializes the given index and search for the k nearest neighbors of the query.

Parameter

type SerializedIndex = string;

type Query = Float32Array; // embeddings of the search query

type NumberOfResult = number; // K top results to return

Return

interface SearchResult {
  neighbors: Array<{
    id: string; // id of the resource
    title: string; // title of the resource
    url: string; // url to the resource
  }>;
}

add(index: SerializedIndex, resource: Resource): SerializedIndex

It adds resources to the index and returns an updated serialized index.

Parameter

type SerializedIndex = string;

interface Resource {
  embeddings: Array<{
    id: string; // id of the resource
    title: string; // title of the resource
    url: string; // url to the resource
    embeddings: number[]; // embeddings of the resource
  }>;
}

Return

type SerializedIndex = string;

remove(index: SerializedIndex, resource: Resource): SerializedIndex

It removes resources from the index and returns an updated serialized index.

Parameter

type SerializedIndex = string;

interface Resource {
  embeddings: Array<{
    id: string; // id of the resource
    title: string; // title of the resource
    url: string; // url to the resource
    embeddings: number[]; // embeddings of the resource
  }>;
}

Return

type SerializedIndex = string;

clear(index: SerializedIndex): SerializedIndex

It removes all items from the index and returns an empty serialized index.

Parameter

type SerializedIndex = string;

Return

type SerializedIndex = string;

Usage

With Transformers

As of now, voy relies on libraries like transformers.js and web-ai to generate embeddings for text:

import { TextModel } from "@visheratin/web-ai";

const { Voy } = await import("voy-search");

const phrases = [
  "That is a very happy Person",
  "That is a Happy Dog",
  "Today is a sunny day",
];
const query = "That is a happy person";

// Create text embeddings
const model = await (await TextModel.create("gtr-t5-quant")).model;
const processed = await Promise.all(phrases.map((q) => model.process(q)));

// Index embeddings with voy
const data = processed.map(({ result }, i) => ({
  id: String(i),
  title: phrases[i],
  url: `/path/${i}`,
  embeddings: result,
}));
const resource = { embeddings: data };
const index = new Voy(resource);

// Perform similarity search for a query embeddings
const q = await model.process(query);
const result = index.search(q.result, 1);

// Display search result
result.neighbors.forEach((result) =>
  console.log(`✨ voy similarity search result: "${result.title}"`)
);

Multiple Indexes

import { TextModel } from "@visheratin/web-ai";

const { Voy } = await import("voy-search");
const phrases = [
  "That is a very happy Person",
  "That is a Happy Dog",
  "Today is a sunny day",
  "Sun flowers are blooming",
];
const model = await (await TextModel.create("gtr-t5-quant")).model;
const processed = await Promise.all(phrases.map((q) => model.process(q)));

const data = processed.map(({ result }, i) => ({
  id: String(i),
  title: phrases[i],
  url: `/path/${i}`,
  embeddings: result,
}));
const resourceA = { embeddings: data.slice(0, 2) };
const resourceB = { embeddings: data.slice(2) };

const indexA = new Voy(resourceA);
const indexB = new Voy(resourceB);

License

Licensed under either of

• Apache License, Version 2.0, (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
• MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT)

at your option.

Sponsor

Contribution

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.