Compare with proprietary models

[IonMich]

We should add an inference endpoint that calls OpenAI's 4o and 4o1 variants, to get an idea of what is the state of the art in DocVQA. Similarly for Anthropic's Claude API and Google's Gemini.

Notably, these models should be able to extract all relevant information from the page in one-shot, including problem statement and handwritten attempted solution.