It is likely that OpenAI's PDF import is using a combination of reading in PlainText as well as doing processing of the page images using its multi modal capabilities. It is also likely that is bringing a number of other programs to bear on the problem. When I use the gpt-4o model on a image of a PDF with equations it does a quite good job (with the correct prompting) to get the content of the page and include the equations with embedded LaTeX strings. As a general rule in work like this, which I've done a bunch of over the past year, if the image of the page has many equations that are themselves quite small, imagine a grid of equations were the equations have very small font sizes, the LLM will often make a number of errors on the content that has a small font size. Also, if there's a fair amount of very complex, two dimensional layout of equations, or different equations laid out next to one another that should be distinguishable from one another, then it will often times make such errors. I have not tested this on open AI's website, all of my work has been within Wolfram language itself. I suppose it's also worth remembering that the import of a PDF using the Import function has all of the restrictions of any other importing such as the various python packages, as far as I understand it. And in Wolfram language, the import function will import plane text and often times, if the text is laid out in a way that is not strictly linear be confused about which pieces of text relate to other pieces of text for example different pieces of text in a grid. Properly prompted vision models will be able, sometimes to make the distinction. But it all depends upon very sophisticated prompting. And for a sufficiently complicated layout of a PDFs page, as well as the level of detail of equations, and so on, will lead to errors.
