/search
Search by meaning. Not keywords.
Every embedding runs on your CPU, in this tab. Nothing gets typed here ever reaches a server: the model is a 5-7 MB WebAssembly file that turns text into a vector, right in the browser, and compares it against our blog and news posts.
Built on ternlight, an open-source on-device embedding engine. We took it apart, understood how it works, and wired our own implementation on top for this site's content, with a live toggle between its two model tiers.
fast, smaller
ASK
QUERY LATENCY
-- ms
single embed() call, @ternlight/mini
MODEL
Transformer layers2
Attention heads4
d_model256
Output384 dims
Max input128 tokens
PACKAGE
Wire size (gzip)5.0 MB
p50 embed latency~2.5 ms
Corpus19 posts
First load--
LicenseMIT