Yes, the whole document via document.documentElement.outerHTML. Surely there's a lot of optimizations we could look into but this was more to try if the model c

Yes, the whole document via document.documentElement.outerHTML. Surely there's a lot of optimizations we could look into but this was more to try if the model could localize elements via XPath by using screenshot + HTML. It's doing a decent job for the most part, I definitely have seen some XPath hallucinations though.