Changing a document schema requires a new extraction test

A supplier revises its form, a customer adds a new field or the startup changes its own target schema. The system still produces JSON, but the downstream process may no longer interpret it correctly.

Version the meaning as well as the field name

Record expected types, permitted values and business definitions. A field called date may mean issue date in one version and service date in another.

Hugging Face’s dataset-feature documentation describes explicit column names and types. Those structures are useful, but business meaning requires additional definition.

An illustrative logistics extractor must distinguish shipment date from expected arrival. Converting both to the same date format does not make them interchangeable.

Replay representative documents

Google Document AI’s evaluation guidance compares predicted entities with labelled test documents. Apply an appropriate evaluation when the extraction target changes, rather than assume prior scores remain valid.

Include old and new formats where both remain in use. Identify whether the process can detect the version or needs a clearer intake rule.

Test downstream validation and exception handling. A schema-compliant value can still violate a business relationship, such as an end date preceding a start date.

Control the release boundary

Record which producer and consumer versions are compatible. Give the receiving system a clear way to reject unsupported output rather than silently guessing.

Add a compatibility test to deployment review that sends the revised output through its actual consumer. A model can satisfy the new schema while the receiving application still expects the old field meaning. The acceptance evidence should therefore include the complete handover, not only successful parsing of the generated object.

Keep corrections and migration assumptions visible to the customer. An extraction change is complete when the operational workflow interprets the values correctly, not merely when the model emits the requested field names.