Invoice prompts: anchor each extracted field

Last updated: 4 Oct 202611 min read

tutorial

Field extraction is a typed proposal, not a posting command. Ask for invoice number, PO reference, quantity, unit price, currency, freight, tax, and total with a page or region anchor. An unreadable digit must be marked uncertain; OCR confidence is not the same as accounting correctness. If two regions disagree, keep both candidates rather than silently preferring the clearer line. The reviewer should be able to inspect every value used in a calculation or matching decision.

Operational case

INV-47 page one shows 47 units at 29. Its tax box is faint, so 95.41 is retained as a candidate with the observed region and a review flag. The printed total is 1,522.41. A parser confirms numeric types and decimal scale before recalculation, but a matched sum does not resolve whether the tax was applied under the correct rule. The model reports unknown where the PDF cannot support a confident field.

INV-47 page 1 line: quantity 47; unit price 29
Summary: freight 64; tax candidate 95.41
Printed total 1,522.41
Each field: type + page region + verified or uncertain

Performance and review cost

Reviewing F candidate fields across P pages is O(P+F) handling after rendering, with human time for uncertain regions. Anchors make a correction local to the disputed value. Without them, an auditor must reread the entire file and cannot distinguish an observed number from a guessed one.

Common Mistakes

Do not post a guessed OCR digit as verified money.
Do not omit field-level page anchors.
Do not flatten conflicting values into one confident answer.

Connected lessons

prompt engineering

invoice operations