In-Browser AI Search for brunosquared.com screenshot

Search my projects and blog posts by meaning instead of by keyword. Type "retro football terminal" and it finds Jornada90, even though that page never uses any of those three words.

The part I actually care about: there is no server. No API key, no backend, nothing sent anywhere. A ~43 MB embedding model gets downloaded once, cached by the browser, and runs in the tab on your own hardware. Whatever you type never leaves your machine.

It uses WebGPU when your browser has it and falls back to WASM when it doesn't. Either way it tells you which one you got, in the panel on the right, along with the actual measured numbers:

  • Which backend it picked, and the reason it picked it
  • Model load and warmup time
  • How long indexing the corpus took
  • How long your last query took
  • Weight precision, switchable between fp32, fp16, q4 and q4f16

Note: the first visit pays for the model download. After that the browser has it cached and the whole thing starts in well under half a second.

The content comes from this site, pulled at build time from a JSON endpoint it publishes. The two are completely separate apps that share nothing but that one URL.


Also in terminal mode

There is a switch in the top left corner. Same search, same model, same numbers on the panel. Just green.

Check it out