Generally I would love to have some bounding boxes come back with the text response. Primarily for highlighting locations in the original document where the text got pulled. Not sure exactly how I would proceed with this one, but would love to hear some thoughts.

I think the general flow would be:

1. Parse the document with gpt mini
2. Split the resulting markdown into semantic sections (i.e. headers, subheaders, tables, etc.)
3. For each semantic section, use [insert ai tool] to find bounding boxes in the original image