Strongly agree! It is basically the Mone-Carlo tree search method used in Alpha Go! This is also mentioned in one of their toturials: PageIndex/blob/main/tutorials/doc-search/semantics.md. I believe it will make the method more scalable for large documents.