Turn multimodal interpretation into editable application data.
1. Output schema
Define fields with clear unknown values and provenance. Keep inferred text separate from confirmed user data and avoid unrestricted generated markup.
2. Cross-field validation
Check units, dates, totals and relationships after parsing. A visually plausible response can violate domain constraints even when it is valid JSON.
3. Correction workflow
Let users compare output with the source media and edit mistakes before saving. Preserve correction data only under the product’s retention policy.
Worked scenario
The assistant reads a serial number with an O instead of a zero. A side-by-side correction flow makes the uncertainty visible.
Apply it
Build a structured field-note form with source preview, validation errors and manual correction.
Check your understanding
Generated fields are not treated as confirmed records before review. Explain the decision and show evidence from your implementation or design. If you cannot demonstrate it yet, revisit the relevant section before continuing.