I’m working on a pipeline to extract structured information from complex PDFs (annual reports, financial filings, technical documents). These PDFs are usually not scanned so they contain real text layers, tables, vector graphics in their meta-data. So OCR is not required.

What I’ve tried so far:

a) PDF Plumber: decent for basic text but breaks when there are tables, images, and complex numbers.
b) Vision models: converting pages to images → better layout understanding but slower and still make mistakes.

What is currently considered the best for extracting high quality structured data from complex PDFs? How are people handling tables reliably and has anyone tried hybrid approaches

I’m processing thousands of long PDFs (100–300 pages each) and care about both accuracy and scalability.

Thank you in advance!