Multiple-page markdown heading level consistency
I haven't tried it yet but I'm very interested. I have a question concerning a key feature for my use-case:
I understand the model takes an image as input (JPEG, PNG). In case of multi-page PDF files of scanned documents, it's easy to extract the corresponding image from each page. But what about the Markdown heading level consistency across pages?
For example if page 1 has a title of level 2, say "## 3.1 Title" and page 2 has a title "3.2 Another title", how can we be sure the model will choose the heading level 2 "##" for it?
I managed to solve this with another model (Claude 3.5 Sonnet) by providing the output of the previous page to the model in order to help it infer the correct level, but it does not work 100%.
If someone has experience on this with this new model, I'm interested.