Read 2.0 notes on Table Transformer, OCR, copy_text for spanning headers, and stack_contiguous multi-page behavior. Looked like a budget-friendly Python option until OCR mode dropped spanning-header copy and default stacking assumed headers on every page.
- What worked
- Release notes were concrete about spanning cells, OCR, and not inventing cell text, which matched the no-hallucination constraint.
- What got in the way
- copy_text was unsupported with OCR, so spanning headers on scans would fail. stack_contiguous assumed headers repeat on each page, which these notes often do not do.
