RAGite
Embeddable, CPU-friendly retrieval (RAG) as a SQLite extension.
1.0 · releases tagged ragite/vX.Y.Z
sqlite> .load ./libragite
sqlite> CREATE VIRTUAL TABLE docs USING ragite;
sqlite> INSERT INTO docs(text, source, format)
...> VALUES (readfile('manual.md'),
...> 'manual.md', 'text/markdown');
sqlite> SELECT chunk, source, score FROM docs_chunks
...> WHERE docs_chunks MATCH 'How do I install it?'
...> LIMIT 5;
-- chunked, embedded on the CPU and indexed on INSERT;
-- MATCH fuses vector similarity and BM25 (RRF)01What it is
RAGite is an embeddable retrieval component for retrieval-augmented generation (RAG). It splits text into chunks, computes embeddings on the CPU and finds the best-matching passages with a hybrid search (vector similarity plus full-text BM25), always with a link back to the source document and position.
It ships as a loadable SQLite extension, so any application in any language that can load SQLite extensions can use it: SQLite for RAG. Documents, chunks, vectors and the full-text index live in the same database file, in process, with no server.
Ask, rank, query
RAGite demo runs hybrid search over the SQLite documentation, 20 Wikipedia articles and these docs, in a single SQLite file.
Ask
A plain question, optionally filtered by source.
Rank
Vectors and BM25 fused; the hit is marked in amber.
Query
Results are rows: SQL, JSON or CSV.
02Quickstart
Download libragite.so from the latest release (Linux x86-64) or build it with cargo build -p ragite-sqlite --release. Start sqlite3 in the directory that holds the file:
SELECT load_extension('./libragite');
SELECT name, technical_id FROM ragite_default_model();
CREATE VIRTUAL TABLE docs USING ragite;
INSERT INTO docs(text, source, format)
VALUES ('# Installation
Run the installer and restart the application.', 'manual.md', 'text/markdown');
SELECT chunk, source, score FROM docs_chunks WHERE docs_chunks MATCH 'How do I install it?' LIMIT 5;Pass the path without .so; SQLite adds the suffix. This block is run as a test, so it stays correct. The user guide covers formats, updates and transactions, linking results to your own tables, the embedding functions and models; its examples run in CI too.
03Interface
| Part | What it does |
|---|---|
CREATE VIRTUAL TABLE docs USING ragite | The Document table. INSERT, UPDATE and DELETE chunk, embed and index the text; ROLLBACK undoes all of it |
docs MATCH 'question' | Hybrid search, one row per Document, ranked by its best chunk |
docs_chunks MATCH 'question' | Hybrid search over chunks, with doc_id, source, start, end, heading_path and score |
| Formats | text/plain, text/markdown, text/html and DoclingDocument JSON |
ragite_embed and friends | Embed text with a table's model, list the models, show the default and a table's model |
Scores come from Reciprocal Rank Fusion (k = 60) of the vector ranking and the BM25 ranking, 50 candidates from each. Chunks hold at most 512 Unicode scalars with a 64-scalar overlap.
04Model
The built-in default model is potion-retrieval-32M, a static Model2Vec model trained for retrieval (English, MIT, 512 dimensions). Its weights are embedded in the extension, so the release libragite.so is about 135 MB. A different model can be built in at build time; each table stores its model's technical id, and a build that does not contain that model refuses the table instead of mixing vectors. A model is not loaded from a path or a URL at run time. No GPU and no ONNX Runtime are needed.
The chunker, the scoring and the other presets are fixed, the same for every caller. Where they live in the code, and how to add a document type, write or change a chunker or change the fusion: docs/tinkering.md.
05Demo
RAGite demo puts RAGite to work: a small Datasette app that searches the SQLite documentation, a few Wikipedia articles and these docs, offline and in one SQLite file.
06License
Licensed under either of MIT or Apache-2.0 at your option (SPDX: MIT OR Apache-2.0). Bundled model weights keep their own license; all third-party licenses are in THIRD_PARTY_NOTICES.md.
07Links
Repository · User guide · Specification · Roadmap · Contributing