Absolutely, it's a crucial part of robust PDF RAG, and I'll definitely test and experiment with tables, multi-column layouts, and extraction-order issues in future iterations. For the current M1 scope, though, I intentionally focused on learning and rigorously evaluating retrieval/reranking rather than PDF extraction complexity.
Kartik N V J K
AI Developer | Making AI reliable, trustworthy & accessible to everyone | Active community contributor
Measuring retrieval and inference separately is exactly the discipline that turns a RAG demo into something you can debug, so I am glad you scoped it that narrowly on purpose. The failure mode I would watch on text-layer PDFs is tables and multi-column layouts, where extraction order scrambles the chunk before retrieval ever runs. Is your reproducible eval set catching extraction errors, or does it assume clean text going in?