Separate extraction from summarization
A PDF may contain selectable text, scanned page images, columns, tables, footnotes, or a mixture of all four. Treating the file as plain text hides extraction errors. A dependable workflow first determines what content can be read, then reconstructs the document’s order, and only then asks for a summary.
Classify the document
- Digital PDF: text can be selected and copied; native extraction is usually the first choice.
- Scanned PDF: pages are images and require optical character recognition.
- Hybrid PDF: some pages contain embedded text while others are scans.
- Layout-heavy PDF: tables, sidebars, captions, and multiple columns require structure-aware extraction.
Run an extraction quality check
Review the first, middle, and last pages. Look for missing headings, merged columns, repeated headers, broken words, page numbers inserted into sentences, and tables flattened into unreadable lines. For screenshots and photos, check skew, glare, cropped edges, small type, and handwriting. Record uncertain pages so the final summary can flag them.
Preserve references that help verification
Keep page numbers, section headings, table titles, figure captions, and document identifiers. When the summary contains a critical statement, attach the nearest page or section reference. For long documents, build a page-range map such as “pages 1–4: background; pages 5–11: findings; pages 12–15: recommendations.”
Summarize tables and figures deliberately
Do not assume a language model will infer a table’s meaning from a flattened text dump. Convert important tables to a simple row-and-column representation, retain units, and identify totals separately from percentages. For figures, provide the caption and a short human description of axes, trends, and visible annotations.
Use a two-stage prompt
- Ask for a document map: purpose, sections, key entities, dates, and evidence locations.
- Ask for the final summary using that map, with unsupported or unreadable points labeled clearly.
This reduces the chance that early pages dominate the output and makes it easier to compare the summary with the source.
Privacy and document handling
Before sending a contract, medical record, financial statement, or internal report to any external service, confirm your organization’s data-handling requirements. Redact secrets and unnecessary personal data. Keep a local copy of the exact extracted text used for the summary so the work can be reproduced and reviewed.
Next steps
For very long files, combine this process with the chapter-mapping method. Developers can use the API guide to design extraction, chunking, and validation as separate pipeline stages.