demo

Two panels, one library

Everything below is computed by moonbitstack/moonrag compiled to JavaScript from the same sources the package ships. There is no second implementation behind these controls: the page imports docs/engine/moonrag.js, built from src/examples/browser/ when the site is published.

Cut a document

Paste anything. The three modes are the three splitters: markdown walks the syntax tree, text tries separators from coarse to fine, and token measures in words — a stand-in for the counter your model really uses.

Code points, not bytes: a Chinese sentence and an English one of the same length cost the same budget. Overlap is capped below the budget, because a chunk cannot carry more of the previous one than it holds.

Rerank a corpus

Seven passages, four numbers each, written down in the engine. Switch between a plain cosine search and MMR, then move lambda: at 1.0 the two orders agree, and as it falls the list starts choosing passages that say something new over passages that say the same thing again.

The score is the cosine to the query. MMR uses it and the similarity between candidates to decide the order; the number shown stays the one a reader can check.

The page cannot drift from the library

The engine is rebuilt from these sources whenever the site is published, so the page cannot drift from the library. A page that reimplemented the splitter in JavaScript would not break; it would quietly disagree, which is worse.

The two things this page cannot show are the ones that need a database: the SQL store and its filters. Those are exercised against SQLite and against PostgreSQL with pgvector, on the same standard cases the in-memory store passes.