It's a shame that it's not open source, unlikely that there's anything super proprietary in an embeddings model that's optimized to run on CPU.
(I'd use it if it were released; in the meantime, MiniLM-L6-v2 works reasonably well. https://brokk.ai/blog/brokk-under-the-hood)
Ideally they would expose the model via a browser api like they do for the prompt api.
FWIW I asked someone on the chrome team about this and they don’t plan to expose native embedding APIs citing lack of dimensionality standards as the reason.
possibly this is buried in the Chromium source if it's part of the OSS personality, if not difficulty of reverse engineering would be pretty high
What it’s for “History Embeddings.” Since ~M‑128 the browser can turn every page‑visit title/snippet and your search queries into dense vectors so it can do semantic history search and surface “answer” chips. The whole thing is gated behind two experiments:
^ response from chatgpt