How do you stop AI from making up facts when reading bad PDF scans?

Tried building a rag app to help people navigate confusing government paperwork like pension files and certificates.

The model handled plain text fine, but started making up entire steps whenever it hit messy or blurry scanned pages.

Reading unstructured pdfs feels like an absolute nightmare right now.

What tools or tricks are you using to clean up terrible documents before feeding them to an ai?