CCARF Practice Question

Paper 1 · Question 17

D4 Prompt Engineering & Structured Output · Paper 1

Your data extraction pipeline uses tool_use with tool_choice set to 'auto'. You need guaranteed structured output for every request, but some documents occasionally produce conversational text responses instead. You have only one extraction tool defined. What is the most reliable fix?

• A.Add a system prompt instruction: 'You must always use the extraction tool and never produce text responses'
• B.Set tool_choice to force the specific extraction tool by name
• C.Add a post-processing step that retries any request returning text instead of a tool call
• D.Switch to prompt-based JSON extraction with strict formatting instructions for more predictable output

Show answer & explanations

Correct answer: B

A: System prompt instructions cannot guarantee tool use when tool_choice is 'auto'. The API-level tool_choice parameter is the authoritative mechanism.

B: Setting tool_choice to {type: 'tool', name: 'extract_data'} forces the model to call that specific tool on every request, guaranteeing schema-compliant structured output with no possibility of text-only responses.

C: Retrying with the same 'auto' tool_choice may produce the same text response. Fixing tool_choice eliminates the problem at the API level rather than patching it after the fact.

D: Prompt-based JSON is less reliable than tool_use. It risks syntax errors and missing fields. Forcing the tool via tool_choice is the correct approach.

Related Prompt Engineering & Structured Output questions

More: Paper 1 · All practice questions