Why Is Table Extraction with VLM Models Still Challenging? [D]
Borderless tables and multi-column PDFs still trip up VLM models
Deep Dive
A Reddit user seeking to convert financial PDFs to Markdown highlights persistent challenges with borderless tables and layouts over 5–6 columns. Open-source tools like docling, graphite-docling, and marker fail; only the paid LandingAI solution works reliably, underscoring a gap for developers needing robust free alternatives.
Key Points
- Borderless tables and layouts with >5–6 columns remain unsupported by open-source VLM tools like docling and marker
- Only the paid LandingAI solution reliably extracts such tables from financial PDFs
- Community seeks better training data or hybrid approaches to bridge the open-source gap
Why It Matters
Financial data extraction remains a pain point for AI workflows due to open-source limitations.