In-browser AI

Real models, downloaded once and run on your GPU (or on WASM when WebGPU is missing). Nothing is pre-computed.

Semantic search

all-MiniLM-L6-v2 (23 MB) embeds ~200 chunks of public Dudoxx documentation once, then ranks them against your query by cosine similarity. No PHI, nothing pre-indexed.

Xenova/all-MiniLM-L6-v2

~23 MB

Model not loaded yet.

Run onactual: —

WASM is the default on purpose: a 23 MB single-pass model is dispatch-bound, so the WebGPU round-trip costs more than the maths. Switch to WebGPU to compare.

Query