finds.dev

kagisearch/vectordb

A minimal Python package for storing and retrieving text using chunking, embeddings, and vector search.

View on GitHubHomepage ↗

★ 801PythonMITupdated Oct 2024

// the take

VectorDB is a small Python package for local text embedding and retrieval: you save text (optionally chunked) and metadata, then query it with plain-text search backed by Faiss or mrpt. It's aimed at developers who want RAG-style retrieval without standing up a vector database service — Kagi uses it internally for exactly that.

It picks the retrieval backend for you based on corpus size — Faiss under ~4000 chunks, mrpt above that — so you don't have to think about index selection. The embeddings performance table (latency vs. MTEB benchmarks) is actually useful for choosing a model, not just marketing filler. You can drop in any HuggingFace sentence-transformers model by name, so you're not locked into their defaults. The API surface is tiny (save/search/clear/dump) and persistence is a single optional file, which makes it trivial to embed in a script or notebook.

Last push was October 2024 — over two years stale as of now, with no sign of active maintenance, so don't expect fixes if a dependency (sentence-transformers, Faiss) breaks compatibility. There's no metadata filtering on search; you get nearest-neighbor over everything you've saved, so multi-tenant or filtered-subset use cases need a workaround on top. Persistence is a flat memory_file — there's no incremental indexing, no concurrent write safety, and no story for datasets that don't fit in memory, so this doesn't scale past single-process, single-user scripts. Swapping the embeddings model after you've already saved data silently produces garbage distances since there's no versioning or migration path for existing vectors.

We read the README, the directory tree and a key source file in November 2026. Nobody pays to be listed. How we judge →

// at a glance

Stars
801
Forks
44
Language
Python
License
MIT
Last commit
1 Oct 2024

// nearby

If vectordb isn't quite it.

All repos →