Docling leads hunt for robust PDF tables

This thread captures a stubborn RAG reality: PDF tables are still not a solved problem, and developers are moving from single-purpose parsers toward hybrid pipelines that mix layout understanding, OCR, and structure-aware export. The practical takeaway is less “find one perfect library” and more “pick the least fragile parser, then store tables in a retrieval-friendly form.”

• –Docling stands out because it combines PDF layout analysis, table structure extraction, OCR support, and export formats like Markdown and lossless JSON in one stack
• –Commenters explicitly say markdown works better than JSON for retrieval because row and column meaning stays readable to embedding models and chunkers
• –For scanned, messy, or multi-column PDFs, the discussion points toward VLM or OCR-first pipelines rather than classic rule-based extractors
• –Broader web research shows newer tools like Marker are pushing table extraction forward with dedicated table converters and optional LLM passes, but even those position the problem as high-accuracy, not perfect-accuracy