Bluesky Answer Search

Answers quoted word for word from the last 100,000 posts of the Artificial (and real) Intelligence list, kept live and searched on your GPU.

Nothing downloaded yet. The first search fetches about 28 MB, which the browser then keeps.

Settings

Posts searched

Sentences in the answer

At most one sentence per post, best match first.

How it works

Ettin-17m turns your query into a vector and your GPU searches the posts for the closest ones. Then Ettin scores every sentence of those posts against the query, and the best sentence from each of the top posts is shown, best match first. Nothing is generated, so nothing can be misquoted.

part	MB over the wire
Ettin-17m embedder, int4 token table, its tokenizer, and the 256-number last layer	10.8
Index: 100,000 posts as 3-bit codes (256 numbers each) plus a scale and bias per post	10.4
The posts' text, gzipped	7.1
The tokenizer library and the WebGPU engines for Ettin and the scan	0.2
total	28.4

Model and engines measured 2026-11-03 with compression on; the index sizes are for 100,000 posts. A GitHub Action rebuilds the snapshot every morning: it asks Bluesky for the list's new posts, embeds them with the same Ettin model, and drops the oldest beyond 100,000. Ettin's last layer is folded from 768 outputs to 256 with no change to any result, and the index keeps 3 bits of each of the 256: against exact search it quotes 9.7 of the same 10 sentences, where the 1-bit index of all 768 it replaced, at the same size, quoted 9.3 (the measurement). This tab then asks Bluesky's public API for the posts written since the snapshot, embeds them on your GPU, adds them to the same index, and checks for new ones every minute while the tab is visible. Your browser keeps a snapshot for two days and fills in the gap itself, so a daily visitor downloads the index every other day. Your query and the results never leave this tab; the only requests after the download are those for new posts.