That Arrow in Every RAG Diagram Cost Us Three Weeks.

Short summary

A developer building a production banking AI chatbot discovered that naive PDF-to-text extraction mangles table structures, producing confidently wrong data in a domain where incorrect numbers have real consequences. After three weeks of trial and error with flat-text and markdown approaches, the team adopted a layout-aware document-intelligence pipeline that preserves table boundaries and stitches together tables split across page breaks. The post is part of a series on building RAG systems for regulated financial domains.

• •Naive PDF extraction destroys table structure, corrupting rate slabs and eligibility data in loan policy documents
• •Markdown table preservation helps but fails on tables spanning page breaks
• •Layout-aware extraction with explicit cross-page table stitching was the fix that worked

Generated with AI, which can make mistakes.