PDF Table Extraction Still Breaks VLMs

The uncomfortable truth is that table extraction is still a full document-understanding problem, not a solved VLM feature. Once you get into borderless layouts, merged cells, and 5+ columns, the failure mode is structural reconstruction, not raw OCR.

• –Open-source tools like Docling, Marker, Camelot, and MinerU each cover part of the stack, but none is a universal fix for messy financial PDFs.
• –The hard part is preserving reading order, row/column boundaries, and cell relationships without turning the result into flattened text.
• –For real-world finance docs, the practical answer is still a hybrid pipeline: layout detection, OCR/VLM fallback, table-structure recovery, and manual review for edge cases.
• –Paid services win here because they ship an opinionated end-to-end workflow instead of exposing parser knobs and hoping users can tune their way out of ambiguity.