Getting usable text and structure out of PDFs, spreadsheets, slide decks and scans. It is dull, it is where most retrieval projects actually lose their quality, and it is almost always underestimated. A table flattened into a wall of numbers is worse than useless downstream.
Technical view
Requires layout analysis to establish reading order, table structure recognition, handling of multi-column text, headers, footnotes and figures, and optical character recognition for scans. Options range from rule-based extractors through layout-aware models to multimodal models reading page images, with the last handling complex layouts best while introducing the risk of plausible invention where text is unclear. Preserving heading hierarchy is what makes structure-aware chunking possible.
Related
Part of AI Terminology Index, section: Retrieval and Knowledge.