I've been thinking lately that maybe we need a new AI-friendly file format rather than continuing to hack on top of PDF's complicated spec. PDF was designed to have consistent and portable page display rendering, it was not a goal for it to be easily parseable afaik, which is why we have to go through these crazy hoops. If you've ever looked at how text is stored internally in PDF this becomes immediately obvious.

I've been toying with an idea of a new format that stores text naturally and captures semantics (e.g. to help with table parsing), but also preserves formatting rules so you can still achieve fairly consistent rendering. This format could be easily converted to PDF, although the opposite conversion would have the regular challenges. The main challenge is distribution of course.

Wouldn’t it be better to invest in a human-friendly format first (which also could be AI-friendly).

If you can convince your bank to make available your bank statement in Markdown, let us know.

Your transactions are probably already available in CSV.

Not really sure what you mean by a "human-friendly" file format, can you elaborate? File formats are inherently not friendly to humans, they are a bag of bytes. But that doesn't mean they can't be better consumed by tools which is what I mean by "AI friendly".

Doesn't Latex do this?

Yea I think Latex is capable of much of this but it's also cursed

Don't need to convince me. I typeset my wife's PhD thesis in LaTeX and it looks great but it was so frustrating that after I did mine in Word.

Sounds like you want XML

Given the current llm context size limitation, what is the state of art for feeding large doc/text blobs into llm for accurate processing?

The current generation of models all support pretty long context now - the Gemini family has had 1m tokens for over a year, GPT-4.1 is 1m, interestingly GPT-5 is back down to 400,000, Claude 4 is 200,000 but there's a mode of Claude Sonnet 4 that can do 1m as well.

The bigger question is how well they perform - there are needle-in-haystack benchmarks that test that, they're mostly scoring quite highly on those now.

https://cloud.google.com/blog/products/ai-machine-learning/t... talks about that for Gemini 1.5.

Here's a couple of relevant leaderboards: https://huggingface.co/spaces/RMT-team/babilong and https://longbench2.github.io/

sorry I should have been more clear, I meant around open source llms. and I guess the question is, how are closed source llm doing it so well. And if OS OpenNote is the best we have...

Mainly I think it's that you need a LOT of VRAM to handle long context - server-class hardware is pretty much a requirement to work with more than ~10,000 tokens.

On my i9 desktop with 128GB RAM and only 8GB VRAM, using llama.cpp I can split the work between both CPU/GPU and get the max 200k context to run on Qwen3 at a decent (human-reading) speed.

Generally use 2.5 flash for this, works incredibly well. So many traditionally hard things can now we solved by stuffing it into a pretty cheap llm haha.

What do you mean by “traditionally hard” in relation to a pdf? Most if not all of the docs I’m tasked with parsing are secured, flattened, and handwritten, which can cause any tool (traditional or ai) to require a confidence score and manual intervention. Also might be that i just get stuck with the edge cases 90% of the time.

Very cool. I'll probably use it, but not for AI. I have lots of pdfs for which an epub doesn't exist.

Or if anything I'll add it to the projects-that-already-do-this-but-havent-yet-found list.

How does it compare to docling?

Docling primarily uses AI models to extract PDF content, this project looks like it uses a custom parser written in Java, built atop veraPDF.

Correct me if I am wrong, but Docling can do both. It has also, among other strategies, a non-AI pipeline to determine the layout (based on qpdf I believe). So these projects are not that different.

While it has a PDF parser, my understanding is that it is mainly used to break a PDF document into chunks, which are then handed off to various specialized models. From its docs: "The main purpose of Docling is to run local models which are not sharing any user data with remote services."

I just tested it on one of my nemeses: PDF bank statements. They're surprisingly tough to work with if you want to get clean, structured transaction data out of them.

The JSON extract actually looks pretty good and seems to produce something usable in one shot, which is very good compared to all the other tools I've tried so far, but I still need to check it more in-depth.

Sharing here in case someone chimes in with "hey, doofus, $magic_project already solves this."

Camelot[1] worked very well for me with bank statements. Disclaimer: I'm one of the core contributors.

[1] https://github.com/camelot-dev/camelot

For 'zoned' extraction, Cermine[0] may be of use as a pre-processing step. Mileage may vary as its tailored towards papers.

[0]: http://cermine.ceon.pl/about.html

There is also Unstract open-source. Structured data extraction + ETL. https://github.com/Zipstack/unstract

I got excited until I read that it was Java/Python based.

I'm looking for a library that can extract data tables from PDF and can be called from a C++ program (for https://www.easydatatransform.com). If anyone can suggest something, I'm all ears.

What makes Java/Python not able to be called from C++, or did you mean you have other requirements that make the project unsuitable?

I can fire up a Java program in a separate process. But it is slow and passing data backwards and forwards is clunky. Much better to be able to do it all in one process.

Just finished migrating to it to replace pdf2docx in a new project I’ve been working on and it is so much better. Thanks for open sourcing OpenDataLoader-PDF!