How to Stop PDF Parsers from Hallucinating Tables out of Thin Air

Short summary

PDF parsers commonly hallucinate tables by treating decorative underlines as borders, resulting in jammed text and phantom table detection. This article explains a context-aware extraction pipeline using proximity math to distinguish borders from formatting, classify regions, scope text to bounding boxes, and eliminate false positives—producing clean, semantic HTML from complex PDFs using vanilla JS and pdfjs-dist.

• •Standard PDF parsers hallucinate tables from decorative underlines and dividers because they process documents sequentially without spatial context
• •Solution: context-aware classifier tags regions as TABLE/PARAGRAPH/HEADING/LIST, uses proximity math (0-5px threshold) to distinguish underlines from borders, and scopes text to bounding boxes
• •Result: deterministic, semantic HTML output without AI hallucination; achieves 99% phantom-table elimination using vanilla JS and pdfjs-dist

Generated with AI, which can make mistakes.