Schema-first
Define JSON schema. Model fills fields. Missing/uncertain → null. Constrained decoding guarantees valid JSON.
Advertisement
Nested structures
Address = {street, city, zip}. Product line = list of items. Nest schemas. Model handles arbitrary depth.
Advertisement
Confidence
Optionally add confidence field per extracted value. Model self-assesses. Post-processing threshold rejects low-confidence extractions.
Multi-page docs
Chunk-then-merge. Extract per chunk. Merge with dedup + conflict resolution. Or long-context extraction if it fits.