Structured output
Structured output is an AI response returned in a fixed, machine-readable layout with named fields and types, instead of free-form paragraphs.
What it means
Software cannot reliably use a paragraph like "the bill seems to be about forty thousand". It needs fields: supplier, bill number, date, total. Structured output asks the model to fill exactly that layout, often defined by a schema, and the system rejects anything that does not fit. This turns reading into a data-entry step that can be validated and posted, with unclear fields marked blank instead of guessed.
Example
An agent reads a handwritten purchase slip at a rice mill and returns fields: party name, bags 120, rate ₹2,150, amount ₹2,58,000. The system checks that 120 times 2,150 equals the amount.
Records and checks
- Record the layout, its version and each rejected output.
- Verify that required fields are present and that totals agree with line items.
- The accounts head reviews rejected records and fields often left blank.
Common mistakes
- Allowing the model to fill a field with a guess when the value is not in the document.
- Never testing messy inputs such as photos, so format errors appear only in live use.
Did not find your answer?Message the lamp team on WhatsApp. Ask about setup, your records, or anything we have not covered here.Ask us on WhatsApp →