I built a small MIT-licensed search lab that runs MiniLM feature extraction in
the browser with Transformers.js. The model revision and library version are
pinned, queries stay on-device, and the 20-product catalogue can be replaced
without a backend or API key.
The larger repository contains an Amazon ESCI retrieve-and-rank reference and a
TREC run preflight for CI. The browser lab is deliberately labelled as a
general-purpose MiniLM demonstration, not an ESCI-trained benchmark.
I would value first-query and cached-query latency, plus the least relevant top
result for one example search. Browser/model details make the report useful.