The idea
Document extraction turns source content into named fields. Define types, allowed formats and missing-value behavior before using a model or OCR tool. Keep the source location for important values. A field may be present but unreadable; distinguish that from a field that was never supplied.
Worked example
A fictional invoice includes invoice number, currency and grand total, but no due date. The extracted due date should remain unknown. A model must not infer it from a standard payment policy unless that inference is explicitly separate from the extracted source fields.
Try it
Draw a fictional invoice and a schema for five fields. Specify a missing-value representation and the supporting source line for each field. Add one unreadable field and write the review status it should receive instead of a fabricated value.