Thanks for the feedback, Joe. We'll look into what's wrong with parsing the React/SPA websites. And maybe do some post-processing for readability. Improved model is in the works too--so hope you give us a try again in a week or so.
I also experienced problems asking it to parse a pdf from researchgate.com, but parsing the same pdf from arxiv.org worked fine. I think it would be a good idea to provide some kind of feedback to the user, what exactly was parsed and used as a prompt. ( I was ready to discard the whole thing as stupid, after the first couple of "I don't know" answers)