$ memista

A RAG chunk store

memista is a natural chunk store for RAG: each chunk carries its embedding, text, and metadata, and search returns all three ranked by similarity. One binary, two files, best under ~100k chunks.

For: Builders of RAG pipelines that don't need a cluster

Retrieval-augmented generation needs a store that holds chunk text, some metadata (source, offsets, tags), and a vector — and hands all three back at query time. That’s exactly memista’s chunk shape.

Why it fits

  • A chunk is { embedding, text, metadata }. Insert them, and search returns the ranked text and metadata directly — no second lookup.
  • Metadata is a string you control (typically JSON), stored and returned verbatim, so you can carry source ids, document offsets, or tags.
  • Isolation is per database_id, so one memista instance can back several corpora, each its own two files.

Where it stops

memista is experimental and tested under ~100k chunks. For a personal or per-project RAG corpus that’s ample; for a multi-tenant SaaS indexing millions of chunks, reach for a distributed store. Compare memista vs pgvector if your metadata already lives in Postgres.

Related: run a nearest-neighbour search, how it works.

Frequently Asked Questions

Does memista store the chunk text, or just the vector?

Both. The text and a metadata string live in SQLite; the vector lives in USearch. Search returns the text and metadata, not just an id.

Related

Try memista

A single crate, GPL-3.0. Two files on disk and three endpoints.