RAGite demo
Search the SQLite docs and Wikipedia in your browser with RAGite and Datasette, offline and in one SQLite file.
demo · uses the RAGite 1.0.1 release · Linux x86-64
$ datasette serve app --load-extension ./libragite
SELECT i.title, c.score
FROM docs_chunks c JOIN doc_info i ON i.doc_id = c.doc_id
WHERE docs_chunks MATCH 'How do I copy a database while it is in use?'
LIMIT 1;
-- top hit: SQLite Backup API, score 0.016401What it is
RAGite demo is a small Datasette app that shows RAGite at work. It indexes the SQLite documentation, 20 Wikipedia articles and RAGite's own docs in one SQLite file and serves a search page next to Datasette's full SQL interface. Nothing runs on a server or calls an external service: the embedding model is built into the extension and runs on the CPU.
Type a question, optionally pick a source, and get passages with the document title, the heading path, the source link and a score. A very short hit, usually a bare heading, comes with the text that follows it, and each passage is shown once. Every search links to the same query as SQL.
From question to passage, in one SQLite file
Four steps through the demo app. Every screen runs on Datasette and the RAGite extension.
Ask in plain words
Type a question and narrow it to a source if you like. No keywords, no query syntax.
Get the best passage first
Vector similarity and BM25 are fused into one ranking. Matching terms are marked in amber.
It is just rows
The same search as a canned query: title, URL, source, chunk. Export as JSON or CSV.
Under the hood: one MATCH
Documents and chunks are SQLite virtual tables. Query them from any SQLite client that loads the extension.
select rowid, doc_id, chunk, score from docs_chunks where docs_chunks match :q limit 20;
02Quickstart
git clone https://github.com/pihme/ragite-demo.git
cd ragite-demo
python3 -m venv .venv && . .venv/bin/activate
pip install -r requirements.txt
scripts/build-demo-db # downloads libragite.so and the pinned sources, builds app/demo.db
datasette serve app --load-extension ./libragite --port 8001
# open http://127.0.0.1:8001/search03Search with SQL
The demo uses only RAGite's public interface: the Document table, its _chunks companion and the ragite_* functions. The examples/ folder has runnable SQL for creating a table, inserting each format, searching documents and chunks, and embedding text.
SELECT i.title, c.chunk, c.score
FROM docs_chunks c JOIN doc_info i ON i.doc_id = c.doc_id
WHERE docs_chunks MATCH 'How do I change the output mode of the sqlite3 shell?'
LIMIT 5;04Data
| Collection | Documents | License |
|---|---|---|
| SQLite documentation 3.54.1 (subset) | 74 | Public domain |
| Wikipedia: SQLite, SQL, B-tree, RAG, BM25, embeddings, vector search and related topics (en, de) | 20 | CC BY-SA 4.0 |
| RAGite README, SPEC, GLOSSARY, user guide, tinkering guide | 5 | MIT OR Apache-2.0 |
99 documents, 8,789 chunks in a 25 MB demo.db, rebuilt from pinned inputs by the build script in a few seconds and not committed. Table rows and lists of short items are joined into paragraphs before RAGite cuts them. Sources, revisions and attribution: DATA_SOURCES.md.
05License
The code is MIT. The indexed texts keep their own licenses (DATA_SOURCES.md, THIRD_PARTY_NOTICES.md).