Extraction Prompts

Best practices for writing effective prompts to extract structured data in fileAI

1. Be Explicit and Specific

  • Instead of "Extract important details", use "Extract the invoice_number, total_amount, and due_date from the text."

2. Match Field Names Exactly

  • Use field names as defined in your schema, e.g., customer_name, not name or client.

3. Use Clear Instructions

  • Example: "Extract the delivery_status_enum as one of: shipped, pending, delayed."

4. Set Expectations for Format

  • For dates: "Extract received_date in the format YYYY-MM-DDTHH:MM:SSZ."
  • For yes/no: "Is the payment confirmed? Respond with true or false."

5. Provide Context When Necessary

  • If the field is ambiguous, clarify in the prompt
  • Example:"Extract the contract_date_time (the date the contract was signed, not created)."

6. Use Examples (if supported)

  • If your system supports few-shot prompting, include examples of expected input/output.

What’s Next