Check more than the first page
Sample a prose page, a diagram, a table and any equation-heavy section. Look for rotation, cropped margins, low contrast and missing pages. Test whether text can be selected, but remember that selectable text can still contain OCR errors. Keep the original scan available for comparison.
Verify the meaning of extracted content
Compare a paragraph and one structured item with the image. Watch for two-column text being interleaved, table headers detached from values and superscripts flattened into ordinary numbers. Correct or exclude damaged passages before asking a tool to generate answers from them.
Generate a small, checked sample first
Start with one section and a few questions rather than the entire document. Verify each answer and source location against the scan. If the extraction fails systematically, use a clearer source or a corrected transcript; repeatedly regenerating a summary cannot restore information that was never extracted.
Worked example
A scanned table lists conditions in columns, but extraction places all values in one line. A generated question associates the wrong value with a condition. The student reconstructs the table with labelled columns and checks a second question before continuing. This avoids turning a layout error into a memorized course fact.
Before you finish
- Text, tables and equations have each been sampled where present.
- Extracted reading order matches the page image.
- A small generated sample is checked before processing the whole file.