Embedding search is effectively machine translation into a single common ‘language’ - embedding space - and then searching that; cleaner and less lossy than translating everything into English for searching, but harder to debug when it goes wrong.