I built a vector database that runs entirely in WebAssembly. No server, no API calls, no data leaves the browser.
Why? Most vector DBs require a server. If you're building privacy-sensitive apps (medical notes, legal docs, personal journals) or offline-first tools, that's a problem. EdgeVec solves it.
What it does:
- Semantic search on embeddings from OpenAI, Cohere, HuggingFace, etc.
- Binary quantization: 32x memory reduction (1M vectors in ~125MB)
- SIMD acceleration: 8.75x faster Hamming distance
- IndexedDB persistence: index survives page reloads
- Filter expressions: SQL-like metadata queries
Technical details:
- 220KB gzipped, 494KB raw
- Chrome 91+, Firefox 89+, Safari 16.4+
- Brute force search (HNSW coming in v0.8)
- MIT licensed
matteo1782•1mo ago
Why? Most vector DBs require a server. If you're building privacy-sensitive apps (medical notes, legal docs, personal journals) or offline-first tools, that's a problem. EdgeVec solves it.
What it does: - Semantic search on embeddings from OpenAI, Cohere, HuggingFace, etc. - Binary quantization: 32x memory reduction (1M vectors in ~125MB) - SIMD acceleration: 8.75x faster Hamming distance - IndexedDB persistence: index survives page reloads - Filter expressions: SQL-like metadata queries
Technical details: - 220KB gzipped, 494KB raw - Chrome 91+, Firefox 89+, Safari 16.4+ - Brute force search (HNSW coming in v0.8) - MIT licensed
Live demo: https://matte1782.github.io/edgevec/demo/
This release includes our first community contribution – @jsonMartin implemented the SIMD Hamming kernel that delivered 8.75x speedup.
Happy to discuss architecture decisions, WASM/SIMD tradeoffs, or binary quantization details.