CCARF Practice Question

Paper 4 · Question 22

D4 Prompt Engineering & Structured Output · Paper 4

The moderation system's classification schema has a 'category' field defined as a free-text string. Auditors find 47 different category values in production data, including 'hate speech', 'Hate Speech', 'hate-speech', 'hateful content', and 'hate_speech' — all intended to be the same category. This makes downstream analytics and routing unreliable. What is the best schema fix?

• A.Add detailed instructions to the prompt listing the exact category names and their capitalisation so the model always emits the canonical string.
• B.Change 'category' from free-text to an enum with values like 'hate_speech', 'spam', and 'harassment', plus an 'other' option.
• C.Add few-shot examples showing the correct category formatting for each type
• D.Add a post-processing normalisation step that maps all variations to canonical category names

Show answer & explanations

Correct answer: B

A: Prompt instructions are probabilistic. The model may still produce variations despite instructions. Schema-level enforcement via enums is deterministic and cannot be overridden.

B: Enum fields constrain the model to predefined values, eliminating spelling and formatting variations. Including 'other' as an enum value handles edge cases without allowing free-text drift. With strict: true, the schema guarantees only valid enum values are returned.

C: Few-shot examples improve consistency but cannot guarantee exact string matching. Schema enums are the correct mechanism for constraining categorical output to a fixed set of values.

D: Post-processing normalisation adds complexity and requires constant maintenance as new variations appear. The schema should prevent the problem at the source rather than patching it downstream.

Related Prompt Engineering & Structured Output questions

More: Paper 4 · All practice questions