PDF Parsing That Didn’t Give Up, Even on “Fax Era” Scans

In DocuChat, we implemented PDF parsing with structured text extraction: we don’t just pull out characters, we preserve the document’s logic — headings, lists, tables. And this is where things get interesting: a normal PDF can still be tamed, but a low-quality scan is already a “survive at any cost” situation.

That’s why we added OCR for scanned documents: we recognize text where the PDF is essentially just an image. After that comes the magic, but the engineering kind: in the chat with AI, the model suggests edits, and the user accepts or rejects changes through a diff interface. Almost like code review, but for documents.

Against the backdrop of the hype around AI assistants, one thing is especially clear: just “reading a PDF” is no longer enough. You need a controlled workflow where a human stays in the loop. That’s exactly what we’re digging into.