i'm using an embedding model on the cluster right now. But if i send the embedding directly (hardcoded) then it should be almost as fast as the keyword-only search?