CCARF Practice Question

Paper 3 · Question 55

D4 Prompt Engineering & Structured Output · Paper 3

[Scenario 6: Structured Data Extraction] You are building a structured data extraction system: pulling information from unstructured documents, validating output against JSON schemas, handling edge cases gracefully, and integrating with downstream systems. An extraction fails Pydantic validation with two specific field errors. You will retry. What should the retry request contain — and when would you skip retrying altogether?

• A.Just the instruction “try again more carefully”; skip retries on Mondays
• B.Only the validation errors, to keep the retry cheap; always retry at least five times
• C.The original document, the failed extraction, and the specific validation errors — enabling targeted self-correction; skip retrying when the required information is simply absent from the source, since no retry can produce it
• D.A different document, to give the model a fresh start

Show answer & explanations

Correct answer: C

A: retries blind — the model can't fix errors it isn't shown

B: removes the document and extraction the correction must reference

C: Retry-with-error-feedback works because the model sees what it produced, what was wrong, and the source to correct against. And the boundary matters: format and structural errors are retryable; information absent from the document is not.

D: abandons the failed extraction rather than repairing it

Related Prompt Engineering & Structured Output questions

More: Paper 3 · All practice questions