I have a fairly broad spread of article types, so I don't honestly expect my implementation to work with 100% accuracy. I'm yet to see what improvements have been provided in the recent update. I'll keep you apprised if I find a solution to my problems with fit alone.

1. Issues in the placement of figure captions below the figures
2. Fails most of the time with multicolumn PDFs
3. Fails with complex infographics
4. Fails with equations

That said, for a regular single-column research paper, it's very good.

Also, PyMuPDF got some remarkable upgrades (in case you haven't seen the docs recently)
pymupdf.readthedocs.io/en/latest/

Hey @jamiecummins.bsky.social, your approach has largely solved the issue of format preservation, although about 25-30% of content is still poorly extracted.
Still working on the other pet project - DocOwl.

I concur. For academic papers, GROBID is the best, although it falters with metadata occasionally.
I'm currently trying out something completely different:
github.com/X-PLUG/mPLUG...

I'll first try your recommended approach. If that works for fitz, it'll save me a ton of time and new explorations.

I'd used fitz as is without any coordinate matching. I'll text it out and share my findings here. Fitz is otherwise superb when it comes to processing speed and overall extraction quality.

Still perusing your blog as we speak. My horrors with fitz made me ask that right off the bat. Thank you kindly for the reply. I'll try out your approach, Jamie.

@jamiecummins.bsky.social a question though: is the text structure maintained well in the raw text outputs you get from pymupdf?
I've found that it ruins the structure of research documents.