Extraction and schemas

Extraction schemas describe the fields to pull from a document -- string, number, boolean, date, currency, enum, array, nested object, document reference, or table -- each configured with a description, an example, whether it's required, a confidence threshold, a PII classification, and whether a source citation is mandatory.

Build visually or in JSON Schema

The schema builder offers a visual field editor for day-to-day use and a live JSON Schema view for anyone who wants to author or review the raw schema directly -- both stay in sync.

Every value, cited

Extraction runs against a schema version and return, per field: the extracted value, a confidence score, and the exact source text it was drawn from -- or an explicit “not found” rather than a guessed value. Nothing is invented.

Human review, without losing the original

Low-confidence or missing-but-required fields surface in a review queue. A reviewer can verify, correct, or reject a value -- the model's original output is never overwritten, only annotated, so the audit trail stays intact.