hanabi1224/RuAnnoy

Rust port of annoy (https://github.com/spotify/annoy)

RuAnnoy

This library is a rust port of spotify/annoy , currently only index serving is supported.

A live demo using web assembly is available at https://annoy-web-demo.vercel.app/

It also provides FFI bindings for jvm, dotnet and dart

Metric	Serve	Build	jvm binding	dotnet binding	dart binding	WASM support
Angular	✅	❌	✅	✅	✅	✅
Euclidean	✅	❌	✅	✅	✅	✅
Manhattan	✅	❌	✅	✅	✅	✅
Dot	✅	❌	✅	✅	✅	✅
Hamming	❌	❌	❌	❌	❌	❌

Install via crates.io

# Cargo.toml
[dependencies]
annoy-rs = "0.1"

Usage

use annoy_rs::*;

let index = AnnoyIndex::load(10, "index.ann", IndexType::Angular).unwrap();
let v0 = index.get_item_vector(0);
let nearest = index.get_nearest(v0.as_ref(), 5, -1, true);

SIMD support

SIMD is supported via std::simd on nightly rust. Note that avx intrinsics need to be enabled explicitly by setting your cpu features in RUSTFLAGS environment variable.

RUSTFLAGS="-Ctarget-feature=+avx" cargo +nightly build --release
# or
RUSTFLAGS="-Ctarget-cpu=native" cargo +nightly build --release

WASM support

Install wasm-pack

wasm-pack build
wasm-pack test --node

simd128 is supported in chrome by default.

To enable simd128, build with below command

RUSTFLAGS="-Ctarget-feature=+simd128" cargo +nightly build --release --target wasm32-unknown-unknown

An example site is deployed at https://annoy-web-demo.vercel.app/

Source code is under example/web

FFI support

kotlin/java

It uses JNI bindings to rust crate and is ~5-10x faster than pure java implementation in benchmark scenario

Note that the prebuilt dynamically linked libraries are built with simd support, avx cpu feature is required.

Install via jitpack.io

repositories {
  mavenCentral()
  maven { url 'https://jitpack.io' }
}

dependencies {
  implementation 'com.github.hanabi1224:RuAnnoy:<tag>'
}

Usage

val index = AnnoyIndex.tryLoad("index.5d.ann", 5, IndexType.Angular)

dotnet

Runtimes	Nuget package
RuAnnoy
RuAnnoy-Batteries-Windows-x64
RuAnnoy-Batteries-Linux-x64
RuAnnoy-Batteries-Darwin-x64

Install via nuget

  <ItemGroup>
    <PackageReference Include="RuAnnoy" Version="*" />
    <PackageReference Include="RuAnnoy-Batteries-Windows-x64" Version="*" />
  </ItemGroup>

Usage

var index = AnnoyIndex.Load("index.5d.ann", 5, IndexType.Angular);

dart

Install via pub.dev

# pubspec.yaml
dependencies:
  dart_native_annoy: ^0.1.0

Usage

import 'dart:ffi';
import 'package:dart_native_annoy/annoy.dart';

/// Creat factory from DynamicLibrary
final indexFactory = AnnoyIndexFactory(lib: DynamicLibrary.open('libannoy_rs_ffi.so'));

/// Load index
final index = indexFactory.loadIndex(
      'index.euclidean.5d.ann', 5, IndexType.Euclidean)!;

print('size: ${index.size}');

final v3 = index.getItemVector(3);

final nearest = index.getNearest(v0, 5, includeDistance: true);

TODO

• Index building support
• CLI tool to build index from file