The quick read
- An invoice extractor may return a valid date string that refers to the due date rather than the issue date.
- Allow a missing or unknown value when the source does not contain an answer.
The output contract
A schema can require fields, constrain types and describe allowed values. Instead of extracting an answer from prose, software can read a defined object. That is useful for document extraction, routing and other workflows where the next step expects specific information.
A shape can still contain an error
An invoice extractor may return a valid date string that refers to the due date rather than the issue date. A category can be allowed by the schema and still be the wrong category. Treat schema validation as the first check, followed by checks against the source and business rules.
Make uncertainty representable
Allow a missing or unknown value when the source does not contain an answer. Keep evidence references where practical. For an illustrative extraction pilot, compare several fields against a manually checked sample and record both omissions and confident errors. Version the schema when requirements change so downstream software can adapt. The aim is a reliable contract between components, not the elimination of all judgement simply because an answer can be parsed.
Sources & notes
AI-assisted editorial content checked against the linked sources.
Sources reviewed for the September 2026 launch edition.