The paper mentions they used Wikipedia as search corpus. The repo states they plan to expand to Google, Bing APIs. I wonder how they will handle evolving search corpora, ie. if continual RL updates will be needed.