Crate vector_lite

VectorLite

The SQLite of Vector Database in Rust.

Features

• Compact size
• Zero-copy de/serialization from/to disk
• Excellent search performance
• WASM support
• Minimal dependencies

Planned

• Concurrent insert/delete/search
• Support usage as external index for Parquet dataset.

Usage

use vector_lite::{VectorLite, Vector, ANNIndexOwned};

// Create a new index with 4 trees and leaf size of 10
const DIM: usize = 3;
let mut index = VectorLite::<DIM>::new(4, 10);
index.insert([1.0, 0.0, 0.0].into(), "101".to_string());
index.insert([0.0, 1.0, 0.0].into(), "102".to_string());
index.insert([0.0, 0.0, 1.0].into(), "103".to_string());
index.insert([0.7, 0.7, 0.0].into(), "104".to_string());

let query = [0.9, 0.1, 0.0].into();
let results = index.search(&query, 2);

for (id, distance) in results {
    println!("ID: {}, Distance: {}", id, distance);
}

index.delete_by_id("102".to_string());

// De/serialize from/to disk
let serialized = index.to_bytes();
let loaded_index = VectorLite::<DIM>::from_bytes(&serialized);

Benchmark

cargo bench --bench ann_benchmark "ANN Index Building/10000"

WASM

Need to build with wasm_js backend for getrandom crate.

env RUSTFLAGS='--cfg getrandom_backend="wasm_js"' cargo build --target wasm32-unknown-unknown

Notable users

VectorLite is the primary vector database used in Seen – knowledge management for the impatient.

License

Apache-2.0

Structs

LinearSearchExternal

A linear search implementation of the ANNIndex trait. This performs no indexing and simply scans the entire dataset for each query. It’s useful as a baseline comparison and for small datasets.

LshExternal

A simple LSH-based ANNIndex implementation.

Vector

VectorLite

A lightweight lsh-based ann index.

VectorLiteIndex

Enums

ScoreMetric

Traits

ANNIndexExternal

A trait for ANNIndex that references external data. The index is read-only once built.

ANNIndexOwned

VectorKey