The SQLite vector extension that recommends you use something else

What it does

sqlite-vss is a SQLite extension that bolts Meta’s Faiss library onto the database engine, exposing vector similarity search through a vss0 virtual table interface. You store embeddings as JSON or raw bytes, then run k-nearest-neighbor queries via a vss_search function in ordinary WHERE clauses. It is strictly bring-your-own-vectors: you generate embeddings elsewhere and hand them to SQLite.

The interesting bit

The extension mirrors SQLite’s own fts5 full-text search API, which means vector columns feel like native citizens rather than foreign blobs. You can tune the underlying Faiss index per column with factory strings—swap the default flat index for an inverted-file index if you want faster queries, though you will pay a steep training tax. The author has since stopped active development and now points users to sqlite-vec, a successor designed to be easier to install and use.

Key highlights

• Bring-your-own-vectors: works with OpenAI, HuggingFace, sentence-transformers, or any other embedding source.
• Virtual table API modeled on SQLite’s native fts5 full-text search extension.
• Per-column Faiss factory strings let you trade indexing time for query speed.
• Bindings for Python, Node.js, Rust, Go, Ruby, Elixir, Deno, and Datasette.
• Runs fully offline inside SQLite.

Caveats

• Not in active development; the author recommends sqlite-vec for new projects.
• Underlying Faiss indices are capped at 1 GB.
• UPDATE operations are unsupported, and small INSERT/DELETE batches are slow.

Verdict

A useful reference if you need Faiss-specific indexing inside SQLite today, but new projects should look at sqlite-vec instead. Skip it if you need mutable vectors or indices larger than 1 GB.