What it does Voy is a k-d tree-based vector similarity search engine compiled to WebAssembly from Rust. You feed it embeddings (from whatever model you like), and it returns nearest neighbors in the browser or on a CDN edge. The whole thing ships at 75KB gzipped, 69KB brotli — smaller than most hero images.

The interesting bit Most vector databases assume a server. Voy assumes the opposite: serialize the index to a string, ship it to a worker, deserialize and search. The API is deliberately thin — class methods or standalone functions, your choice — because the point is portability, not kitchen-sink features.

Key highlights

• 75KB gzipped / 69KB brotli, tree-shakeable for dead-code elimination
• k-d tree indexing with add, remove, clear, and serialize/deserialize lifecycle
• Designed for Web Workers and CDN edge runtimes (Cloudflare, etc.)
• TypeScript types included, though the README notes WASM tooling limits auto-generation of complex types
• Dual MIT/Apache-2.0 licensed

Caveats

• API is explicitly unstable pre-1.0; breaking changes expected
• No built-in embedding generation — you bring your own transformers.js, web-ai, or equivalent
• The “fast” claim in the README is aspirational; no benchmarks or latency numbers are provided

Verdict Worth a look if you’re building client-side semantic search, offline-capable tools, or edge-hosted recommendations and want to avoid a round-trip to a vector database. Skip it if you need hosted managed scaling, built-in model serving, or a stable API today.