Markdown for RAG - Can't Get Tables for Core Filings Right

[NamedIdentity]

I've been working with GPT 5 to write scripts. I can get the filings I want from years 2014-2025 for the companies I want downloaded, but when I try markdown, problems.

At first It was hard to get anything other than the default primary report. Eventually I got the exhibits I needed to markdown too. I even got forms 3, 4, 5 xmls through markdown, alongside the html based filings, all in the same script.

But, no matter what I try, I can't get the tables in core filings to output in a structured manner. They're always jumbled. The more I try to get GPT 5 to fix, the more I go around in circles and find new ways to break things (at one point data from tables displayed as "nan").

The kicker. If I open the html of a 10-K, print PDF, upload the PDF to Abacus.ai database, it'll extract into markdown with properly structured tables. So, it can be done. I just have no clue how.

I keep thinking, someone, somewhere, has already made the script I need to use, to pull specific forms, clean, parse, and markdown in preparation for RAG.

Which, leads to my request for guidance. What do I need to point GPT 5 to so it figures out how to salvage this script?

Anyone have a script they could share that I can give to GPT as an example of 'what to do'?

I'm 'almost there' for a proper workflow for this. My backup, is to just pull filings and convert to PDF and hope and pray it doesn't create other new problems.