hora

Features

• Performant ⚡️

  • SIMD-Accelerated (packed_simd)
  • Stable algorithm implementation
  • Multiple threads design

• Supports Multiple Languages ☄️

  • Python
  • Javascript
  • Java
  • Go (WIP)
  • Ruby (WIP)
  • Swift (WIP)
  • R (WIP)
  • Julia (WIP)
  • Can also be used as a service

• Supports Multiple Indexes 🚀

  • Hierarchical Navigable Small World Graph Index (HNSWIndex) (details)
  • Satellite System Graph (SSGIndex) (details)
  • Product Quantization Inverted File(PQIVFIndex) (details)
  • Random Projection Tree(RPTIndex) (LSH, WIP)
  • BruteForce (BruteForceIndex) (naive implementation with SIMD)

• Portable 💼

  • Supports WebAssembly
  • Supports Windows, Linux and OS X
  • Supports IOS and Android (WIP)
  • Supports no_std (WIP, partial)
  • No heavy dependencies, such as BLAS

• Reliability 🔒

  • Rust compiler secures all code
  • Memory managed by Rust for all language libraries such as Python's
  • Broad testing coverage

• Supports Multiple Distances 🧮

  • Dot Product Distance
  • Euclidean Distance
  • Manhattan Distance
  • Cosine Similarity

• Productive ⭐

  • Well documented
  • Elegant, simple and easy to learn API

Installation

Rust

in Cargo.toml

[dependencies]
hora = "0.1.1"

Python

$ pip install horapy

Javascript (WebAssembly)

$ npm i horajs

Building from source

$ git clone https://github.com/hora-search/hora
$ cargo build

Benchmarks

by aws t2.medium (CPU: Intel(R) Xeon(R) CPU E5-2686 v4 @ 2.30GHz) more information

Examples

Rust example [more info]

use hora::core::ann_index::ANNIndex;
use rand::{thread_rng, Rng};
use rand_distr::{Distribution, Normal};

pub fn demo() {
    let n = 1000;
    let dimension = 64;

    // make sample points
    let mut samples = Vec::with_capacity(n);
    let normal = Normal::new(0.0, 10.0).unwrap();
    for _i in 0..n {
        let mut sample = Vec::with_capacity(dimension);
        for _j in 0..dimension {
            sample.push(normal.sample(&mut rand::thread_rng()));
        }
        samples.push(sample);
    }

    // init index
    let mut index = hora::index::hnsw_idx::HNSWIndex::<f32, usize>::new(
        dimension,
        &hora::index::hnsw_params::HNSWParams::<f32>::default(),
    );
    for (i, sample) in samples.iter().enumerate().take(n) {
        // add point
        index.add(sample, i).unwrap();
    }
    index.build(hora::core::metrics::Metric::Euclidean).unwrap();

    let mut rng = thread_rng();
    let target: usize = rng.gen_range(0..n);
    // 523 has neighbors: [523, 762, 364, 268, 561, 231, 380, 817, 331, 246]
    println!(
        "{:?} has neighbors: {:?}",
        target,
        index.search(&samples[target], 10) // search for k nearest neighbors
    );
}

thank @vaaaaanquish for this complete pure Rust 🦀 image search example, For more information about this example, you can click Pure Rust な近似最近傍探索ライブラリ hora を用いた画像検索を実装する

Python example [more info]

import numpy as np
from horapy import HNSWIndex

dimension = 50
n = 1000

# init index instance
index = HNSWIndex(dimension, "usize")

samples = np.float32(np.random.rand(n, dimension))
for i in range(0, len(samples)):
    # add node
    index.add(np.float32(samples[i]), i)

index.build("euclidean")  # build index

target = np.random.randint(0, n)
# 410 in Hora ANNIndex <HNSWIndexUsize> (dimension: 50, dtype: usize, max_item: 1000000, n_neigh: 32, n_neigh0: 64, ef_build: 20, ef_search: 500, has_deletion: False)
# has neighbors: [410, 736, 65, 36, 631, 83, 111, 254, 990, 161]
print("{} in {} \nhas neighbors: {}".format(
    target, index, index.search(samples[target], 10)))  # search

JavaScript example [more info]

import * as horajs from "horajs";

const demo = () => {
    const dimension = 50;
    var bf_idx = horajs.BruteForceIndexUsize.new(dimension);
    // var hnsw_idx = horajs.HNSWIndexUsize.new(dimension, 1000000, 32, 64, 20, 500, 16, false);
    for (var i = 0; i < 1000; i++) {
        var feature = [];
        for (var j = 0; j < dimension; j++) {
            feature.push(Math.random());
        }
        bf_idx.add(feature, i); // add point
    }
    bf_idx.build("euclidean"); // build index
    var feature = [];
    for (var j = 0; j < dimension; j++) {
        feature.push(Math.random());
    }
    console.log("bf result", bf_idx.search(feature, 10)); //bf result Uint32Array(10) [704, 113, 358, 835, 408, 379, 117, 414, 808, 826]
}

(async () => {
    await horajs.default();
    await horajs.init_env();
    demo();
})();

Java example [more info]

public void demo() {
    final int dimension = 2;
    final float variance = 2.0f;
    Random fRandom = new Random();

    BruteForceIndex bruteforce_idx = new BruteForceIndex(dimension); // init index instance

    List<float[]> tmp = new ArrayList<>();
    for (int i = 0; i < 5; i++) {
        for (int p = 0; p < 10; p++) {
            float[] features = new float[dimension];
            for (int j = 0; j < dimension; j++) {
                features[j] = getGaussian(fRandom, (float) (i * 10), variance);
            }
            bruteforce_idx.add("bf", features, i * 10 + p); // add point
            tmp.add(features);
          }
    }
    bruteforce_idx.build("bf", "euclidean"); // build index

    int search_index = fRandom.nextInt(tmp.size());
    // nearest neighbor search
    int[] result = bruteforce_idx.search("bf", 10, tmp.get(search_index));
    // [main] INFO com.hora.app.ANNIndexTest  - demo bruteforce_idx[7, 8, 0, 5, 3, 9, 1, 6, 4, 2]
    log.info("demo bruteforce_idx" + Arrays.toString(result));
}

private static float getGaussian(Random fRandom, float aMean, float variance) {
    float r = (float) fRandom.nextGaussian();
    return aMean + r * variance;
}

Roadmap

• [ ] Full test coverage
• [ ] Implement EFANNA algorithm to achieve faster KNN graph building
• [ ] Swift support and iOS/macOS deployment example
• [ ] Support R
• [ ] support mmap

Related Projects and Comparison

• Faiss, Annoy, ScaNN:

  • Hora's implementation is strongly inspired by these libraries.
  • Faiss focuses more on the GPU scenerio, and Hora is lighter than Faiss (no heavy dependencies).
  • Hora expects to support more languages, and everything related to performance will be implemented by Rust🦀.
  • Annoy only supports the LSH (Random Projection) algorithm.
  • ScaNN and Faiss are less user-friendly, (e.g. lack of documentation).
  • Hora is ALL IN RUST 🦀.

• Milvus, Vald, Jina AI

  • Milvus and Vald also support multiple languages, but serve as a service instead of a library
  • Milvus is built upon some libraries such as Faiss, while Hora is a library with all the algorithms implemented itself

Contribute

We appreciate your participation!

We are glad to have you participate, any contributions are welcome, including documentations and tests. You can create a Pull Request or Issue on GitHub, and we will review it as soon as possible.

We use GitHub issues for tracking suggestions and bugs.

Clone the repo

git clone https://github.com/hora-search/hora

Build

cargo build

Test

cargo test --lib

Try the changes

cd examples
cargo run

License

The entire repository is licensed under the Apache License.