Three query embedders, side by side
61,686 posts from the "Artificial (and real) Intelligence" Bluesky list, searched three ways. Each query is embedded in this browser by a hand-written WebGPU engine, and the posts' 2-bit codes are scanned on your GPU. Once everything is cached, the page works with the network off.
What the columns are
Ettin-17m and leaf-ir were tuned on in-domain query/post pairs, and each searches an index of the posts it embedded itself. The EG2 4-layer cut embeds queries into EmbeddingGemma 2's space, and searches the posts as full EG2 embedded them. hit@10 is on 1,498 held-out synthetic queries. Download is the brotli size of the model and tokenizer; each index is another 12.3 MB. Engines: ettin-native/web/engine.js for Ettin and leaf-ir, eg2-wgsl-engine for the cut, onnxruntime-web as the fallback. The cold-launch timings cover reading the model (from the offline cache or the network), engine startup, the first query's embedding, and the scan.