Tables & Formulas

Two elements usually decide whether a parsed document is truly usable: tables and formulas.

Tables: structure first

The most stable goal in Markdown is preserving row × column structure. Visual styling matters less than maintaining semantic alignment.

When tables look broken, first identify the source:

• Text-native tables: usually more stable
• Scanned tables: depend heavily on OCR and layout understanding, and are more error-prone

Formulas: prefer a parseable representation

For downstream systems (rendering, computation, or retrieval), a parseable formula expression is more useful than a purely visual match.

If your goal is retrieval/RAG:

• prioritize not losing formulas
• standardize representations if your pipeline supports it

Practical tips

• sample before setting defaults
• isolate “table-heavy” PDFs into a dedicated strategy
• keep failure examples as regression cases

Related:

• Markdown Output
• Performance & Cost