fileAI fills each schema field by reading its title and description, so the description works as the field’s extraction prompt. A clear, specific description is the most direct way to improve a field that comes back wrong or empty. Before rewriting a prompt, check that the file itself reads well: a digitally generated PDF with selectable text beats a low-resolution scan. See Prepare your files.

​
How it works

Every field in an AI Schema has a title, a type and a description. fileAI sends the description to the extraction model as the instruction for that field; when a field has no description, its title is used instead. You write prompts in two places:
  • In the app: click the sparkle icon next to a field in the AI Schema tab, describe how the field should be captured, then click Run and Save.
  • Over the API: set description on a field with PATCH /v1/files/schema, then re-fill the file with PATCH /v1/files/schema/rerun.

​
Guidelines

Name exactly what you want. Instead of “Extract important details”, write “The invoice number printed in the header, labelled ‘Invoice No.’ or ‘Inv #’”.
When one description mentions another field, use the title as it appears in your schema, such as Customer name, not “name” or “client”.
For a fixed set of answers, list them: “Delivery status: one of shipped, pending, delayed.” Better still, make the field an enum with those values. See Data types.
For dates: “The date received, as YYYY-MM-DD.” For yes/no questions: “Is the payment confirmed? Answer true or false.”
If a document has several candidates, say which one you mean: “Contract date: the date the contract was signed, not the date it was created.”
A short example of the expected output, such as “e.g. INV-2025-0042”, helps the model match the shape you want.

​
Example

AI Schemas

Edit fields in the app or over the API.

Data types

Pick the right type so values come back in the right shape.