vers

Lightweight, simple, single instance, local in-memory vector database written in Rust.

Currently supports the following indexing strategies:

1. IVFFlat (k-means for partitioning)
2. Locality-sensitive hashing (LSH) heavily inspired by fennel.ai's blog post.
3. Hierachical Navigable Small Worlds (HNSW)

This repository is educational. It was meant to understand how ANN search algorithms work under the hood for modern machine learning and is not meant for production (as a cursory scan of the codebase can probably tell you).

"What I cannot build, I do not understand." - Feynman

Getting Started

Like any sensible package, the API aims to be dead simple.

0. Import, obviously:

use vers::indexes::base::{Index,Vector};use vers::indexes::ivfflat::IVFFlatIndex;use vers::indexes::hnsw::HNSWIndex;

1. Build an index:

letmut index = IVFFlatIndex::build_index(
num_clusters,
num_attempts,
max_iterations,&vectors
);// or hnswletmut index = HNSWIndex::build_index(
num_layers,
ef_construction,
ef_search,
num_neighbours,
vectors
)

2. Add an embedding vector into the index:

 index.add(Vector(*emb).normalize(), emb_unique_id);

3. Persist the index to disk:

let _ = index.save_index("wiki.index");

4. Load the index from disk:

// or use HNSWIndex::load_index, ANNIndex::load_indexlet index = matchIVFFlatIndex::load_index("wiki.index"){Ok(index) => index,Err(e) => panic!("Failed to load index! {}", e),};

5. And of course, actually search the index:

let results = hnsw.search_approximate(
embs.get("king"),// query vector10// top_k);// kings, queen, monarch, ...

As shown above, all the indexes share the same API, whether IVFFlat, HNSW or LSH.

Python Bindings (WIP!)

vers now has a simple Python API using pyo3. To use vers with Python (>- 3.7),

importvers# load wiki sample embeddings of dims 300embeddings=vers.load_wiki()
# instantiate the hnsw index with paramshnsw=vers.HNSW(ef_construction=100, num_layers=8, ef_search=32, num_neighbours=8)
# build the indexhnsw.build_index(embeddings)
# search for a query vectorresults=hnsw.search(embeddings.get("king"), top_k=10)

Running adhoc benchmark tests

 cargo build --release
samply record ./target/release/vers