# Check a PDF extraction before asking AI to summarize it A PDF summary can only be as reliable as the text the model receives. Check the extraction against the original pages before interpreting it. Missing rows, merged columns, and OCR errors can change a number or remove the condition that gives it meaning. ## Check the shape of the document Compare the page count and section headings. Look at one page with dense text, one table, and any page containing totals or signatures. If the PDF is scanned, inspect the OCR for names, dates, decimals, and symbols. Do not assume every page uses the same extraction quality. ## A fictional table failure A source table contains “Service fee: $150” and “Total due: $1,050.” Extracted text might place the labels beside the wrong amounts. A summary could then report a $1,050 service fee. The prose may be fluent while the underlying pairing is wrong. | Check | What to compare | | --- | --- | | Rows and columns | Labels remain attached to values | | Decimal marks | 1.50 has not become 150 | | Footnotes | Exceptions remain linked to the relevant row | | Page breaks | Continued tables are not treated as new totals | ## Use AI for extraction questions ```text Review this extracted text for signs of missing pages, merged columns, repeated headers, broken tables, and suspicious numbers. Do not repair uncertain values by guessing. List the original pages I should inspect and the exact issue to check. Document inventory: [page count and sections] Extracted text with page markers: [paste] ``` ## Resolve gaps before summarizing Re-extract a problematic page, manually transcribe the relevant table, or provide a clearly labeled excerpt. Keep the original alongside any correction. If a value remains unreadable, mark it unreadable instead of using the nearest plausible number. Only then ask for a summary that preserves page references and flags missing material. For consequential documents, check every value that affects a decision directly against the PDF. The extraction check is a separate step from evaluating whether the document itself is accurate, current, or authoritative. --- SkillStall · 2026-10-02