hnswlib low-overhead JVM wrapper

JNI bindings for HNSW kNN library Generated using bytedeco javacpp library The intent is provide fast hnsw implmentation for JVM.

Building

gradle init
./gradlew generateLib

Will generate JNI bindings in Java and C and compile them for the current platform.

Usage

Here is example that creates 15k embeddings (size 100), saving them into the index on disk, reading the index and retrieving some distances.

import java.io.File;
import java.io.IOException;

public class HelloHnsw {

    public static void main(String[] args) throws IOException {
        int dimension = 100;
        int nbItems = 15000;
        float seedValue = 0.5f;
        int M = 16;
        int efConstruction = 200;

        HnswIndex index = HnswIndex.create(Metrics.Euclidean, dimension);

        index.initNewIndex(nbItems, M, efConstruction);
        for (int i = 0; i < nbItems; i++) {
            float[] vector = getVector(dimension, seedValue / (i + 1));
            index.addItem(vector, i);
        }

        String path = new File("./bin", "index.hnsw").toString();
        index.save(path);

        index.load(path);
        System.out.println("Loaded" + index.getNbItems() + " items");
        long[] ids = index.getIds();

        for (int i = 0; i < 20; i++) {
            long queryId = ids[i];
            float[] query = index.getItem(queryId);
            KnnResult[] results = index.knnQuery(query, 3);
            for (KnnResult result : results) {
                System.out.println(queryId + " -> " + result.getItem() + " distance: " + result.getDistance());
            }
        }
        index.unload();
    }


    private static float[] getVector(int dimension, float value) {
        float[] vector = new float[dimension];
        for(int i = 0; i < dimension; i++) {
            vector[i] = value;
        }
        return vector;
    }
}

To run same example from repo:

./gradlew run

To run tests:

./gradlew test