// the find
alibaba/zvec
A lightweight, lightning-fast, in-process vector database
Zvec is an embedded vector database from Alibaba, written in C++ with Python, Node, Go, Rust, and Dart bindings. It runs in-process — no server, no separate daemon — and is battle-tested inside Alibaba before being open-sourced. It targets the same niche as ChromaDB or LanceDB but with more serious performance engineering and a broader index menu (HNSW, DiskANN, sparse, hybrid).
The SIMD dispatch layer is thorough — separate AVX/AVX512/AVX512FP16/NEON/SSE implementations for every distance type, not just a single fallback path. DiskANN support means you can index hundreds of millions of vectors without keeping everything in RAM, which is the real scaling wall for in-process databases. Hybrid retrieval combining dense vectors, sparse vectors, and BM25 full-text search in a single query is genuinely useful and something most embedded competitors still lack. WAL-backed durability with multi-process concurrent reads is a legitimate production feature, not an afterthought.
The C++ core is not exposed through a stable ABI — the Python and Node bindings ship pre-built wheels, so if your platform or Python version isn't in the matrix, you're building from source with a non-trivial CMake setup. Writes are single-process exclusive, which is fine for embedded use but means you can't have two services sharing the same collection without a wrapper layer. Documentation is thorough for the Python SDK but thin for Go and Rust, which are listed as 'high-performance' but live in separate repos with their own release cadence. The project is v0.5 and the API surface has changed meaningfully between releases, so collections created with an older version may not open cleanly with a newer one.