boost / docs / try the search

Live · runs entirely in this tab

The ranking boost ships, with nothing installed.

This is the same BM25 engine boost search runs, over a real catalogue of 743 skills, rules and workflows from six registries. The index is fetched once and scored in your browser — no server, no account, no install.

743Items indexed
6Registries tapped
333 KBIndex, gzipped on the wire
0Bytes sent anywhere
Honest limits

What this demo is, and is not

A demo that quietly showed you the good engine and shipped you the other one would not be a demo. So: exactly what runs here, and what does not.

This is boost's keyword engine: BM25 over the full text of every item, with the same k1=1.2 / b=0.75 parameters and the same tokenizer the CLI uses. The scoring runs in your browser against a 1.27 MB index that arrives as 333 KB gzipped — measured on the wire from GitHub Pages, not with local gzip, which compresses harder than a CDN does and would flatter it by 4%.

It is not the semantic engine. boost search can also fuse BM25 with dense vectors, which is what lets “my containers keep restarting” find a Docker skill that never uses the word. Embedding your query needs a ~133 MB model, so a page that pretended to demo it would either hand you that download or bill someone for a hosted endpoint. Measured on the published eval, the gap is real: on natural-language queries BM25 scores hit@1 0.340 against hybrid's 0.440.

So: what you can try here is the floor, and it is the floor every install gets for free. The eval page has the full engine comparison, how chat works covers the grounded answer path, and the guide covers installing the rest.