vekidx

Vector similarity search for Node, written in Rust

• Exact search
• No runtime dependencies
• One file on disk

Install

npm install vekidx

Use

const { VectorIndex } = require("vekidx");

const index = new VectorIndex(384);
// Float32Array, all vectors end to end
index.addBatch(vectors);

// [ { index: 42, score: 0.91 }, ... ]
const hits = index.search(query, 5);
// same, off the main thread
await index.searchAsync(query, 5);

index.save("./my.vidx");
const back = VectorIndex.load("./my.vidx");

Vectors must be normalized to length 1.

Most embedding models give normalized vectors already.

Speed

100k vectors × 384 dims, Apple M-series:

Action	Time
search	12 ms
the same algorithm in JavaScript	57 ms
save (146 MB)	74 ms
load	27 ms

Measured with bench.js in this repository.

Limits

• Brute force Reads every vector on each search (Good to ~1 million)
• One process Not shared between Node worker threads
• search blocks the JavaScript thread Use searchAsync in a server

Build from source

npm install
npm run build
npm run check