Comment by gojomo

With regard to privacy-preservation, there's also potential from the oblivious-retrieval/homomorphic-encryption/zk-proof/etc worlds that one might be able to offer a frozen model/embedding-set, and prove to clients that…

(1) their text was faithfully converted to an embedding according a disclosed technique & stable (but private) set of model-weights; and

(2) here are the top-N neighbors doc-IDs/docs by similarity

…without either the client receiving the full model weights or even their text's precise embedding, nor the server seeing a cleartext version of the client's text or any usable info about what it contains.