How do confidence scores work in document extraction?
Confidence scores estimate how reliable an extracted field is, based on document evidence, model certainty, grounding, validation checks, and schema fit.
Short answer
Confidence scores estimate how reliable an extracted field is, based on document evidence, model certainty, grounding, validation checks, and schema fit.
What this means in practice
Confidence should be field-level, not just document-level. One invoice may have a highly reliable total but a questionable tax ID.
Teams usually define thresholds: auto-approve high-confidence fields, review medium-confidence fields, and reject or reprocess low-confidence fields.