Enterprise knowledge lives in PDFs, spreadsheets, scanned forms and slide decks. If extraction scrambles tables or misreads scans, retrieval and generation inherit the errors.
Practical steps
- Inspect extraction output by eye for a sample of each document type before building anything on top.
- Preserve structure — headings, lists and table cells — rather than flattening to plain text.
- Use OCR for scans, and check accuracy on your own documents, especially numbers and names.
- Handle tables as units, with column headers carried into every row or chunk.
- Flag low-confidence extractions for review.
When answers are wrong, read the extracted text first. Many 'model errors' are extraction errors.