Most RAG tutorials assume clean paragraphs. Real documents — invoices, syllabi, lab manuals — carry tables and figures that a naive text extractor flattens into unreadable rows of numbers.
Extract structure, not just text
Use a layout-aware parser that separates tables from prose and keeps row/column relationships (as markdown tables or JSON), rather than concatenating cell text into one blob. Store each table as its own chunk with a caption describing what it contains.
Images need a caption before they need an embedding
Run a vision-capable model once at ingestion time to generate a text caption of each figure, then index the caption alongside the surrounding paragraph. Retrieval then works off text even though the original content was visual.
- Keep the table's caption or preceding sentence attached to it — a bare table with no context retrieves poorly.
- Store page number and figure number as metadata so citations point somewhere useful.
- Test retrieval with questions that specifically require a table value, not just prose facts.
Related: an OCR pipeline for turning PDFs into searchable text and document AI for invoice extraction.
— Pranjul Rathour, GenAI Engineer from Kanpur, India. Open to GenAI roles, hackathon judging, mentorship sessions and guest talks at any campus: pranjulrathour41@gmail.com.