This is an awesome notebook.

Just want to note that the difference in this paper is that it works without direct access to the embedding models (encoder). So it can't design the embedding space.