CCARF Practice Question

Paper 1 · Question 44

D4 Prompt Engineering & Structured Output · Paper 1

Your extraction pipeline uses tool_use with a JSON schema, so outputs always parse. Finance still reports invoices where line items do not sum to the stated total, and where the due date precedes the invoice date. The team wants retries to receive precise, per-field error messages. What should you add?

• A.Tighter JSON schema constraints: numeric minimum and maximum bounds on every amount field and format checks on every date field
• B.An instruction telling the model to double-check its own extraction in the same request before responding
• C.A Pydantic model whose validators encode the sum and date-order rules, feeding its validation errors into retries
• D.A more capable model, since semantic errors indicate the current one is under-powered

Show answer & explanations

Correct answer: C

A: Range and format constraints check fields in isolation; both reported failures are relationships between fields, which a JSON schema cannot express.

B: Same-session self-review retains the generation context and its bias, and produces no machine-readable error for the retry loop to feed back.

C: Cross-field semantic rules live in validation code; Pydantic validators express them and raise per-field errors the retry request can quote verbatim.

D: No model tier guarantees semantic correctness; the fix is validation logic around the model, and an upgrade is the maximum-cost non-answer.

Related Prompt Engineering & Structured Output questions

More: Paper 1 · All practice questions