RAGEngineering Hub
Menu

Architecture

How this site works

A course about retrieval systems, built so that it needs no server of its own. Everything is computed at build time or in your browser.

Site architecture: at build time, MDX pages, the curriculum CSV and the precompute script produce static HTML, a search index and lab data with vectors. A static host serves these files. In the browser, the labs score questions, embed typed queries with a 34 MB model, and optionally call the model provider directly with the reader's own key.BUILD TIME · ON MY LAPTOPMDX pages + CSV54 ids, 18 writtenCorpus + golden set25 docs, 52 questionsnext buildvalidate, static HTMLnpm run precompute16 settings, vectorsPagefindsearch indexpublic/dataJSON, 3.4 MBStatic hostplain filesRUN TIME · IN THE READER'S BROWSERPages and searchHTML + Pagefind JSLabsBM25, cosine, RRF in JSQuery embedderbge-small, 34 MBModel APIyour key, optionalonly call with a key, direct to the provider
There is no server of mine anywhere in the picture. Running cost is the static host, which is free at this size.

Static export

Next.js builds every page to plain HTML at build time. A content validator fails the build on a broken page, a missing quiz explanation or a prerequisite that does not exist.

Why
No server to run, patch or pay for. Any static host can serve it.
Trade-off
No server features: no accounts, no stored answers. Progress lives in the reader's browser.

Precomputed vectors

npm run precompute chunks the 25 documents in 16 settings, embeds every chunk and every golden question with bge-small-en-v1.5, and writes JSON to public/data.

Why
Labs load the vectors instead of computing them, so a lab opens in about a second, and the numbers on every page can be reproduced with one command.
Trade-off
3.4 MB of lab data in the repository, rebuilt whenever the corpus or chunker changes.

Labs in the browser

BM25, cosine similarity, reciprocal rank fusion and the metrics run as plain TypeScript in the page. A test checks that the browser code gives the same evidence recall as the precompute script for all 16 settings.

Why
Every control changes a real computation, not a canned answer, and nothing leaves the reader's machine.
Trade-off
Only small corpora fit. A real system with millions of chunks needs a vector database.

Typed questions embedded on demand

The Mini RAG Lab loads the same 34 MB quantized model in the browser the first time a reader types a question, with a progress bar. The browser caches it.

Why
Free text search with no API and no cost, in the same vector space as the stored chunks.
Trade-off
A one-time 34 MB download, so it loads only when asked for.

Bring your own key

For a live answer, the reader pastes their own Anthropic key. It stays in session storage (local storage only if they tick remember) and goes only to the provider's API. A test checks the request host and headers.

Why
Live answers with no proxy and no bill for the site.
Trade-off
Readers without a key see the reference answer from the golden set instead.

Try it: the labs are the same code described here.