A claims-support bot answers from retrieved policy excerpts. The team wants responses that drift away from those excerpts to be caught automatically, using a score they can set a threshold on. Which Guardrails feature provides this?
Show answer and explanation
Answer: B
The contextual grounding check scores a model response for grounding, meaning whether it is supported by the reference passages, and for relevance to the user query. Each score is compared with a configurable threshold between 0 and 1, and responses below it can be blocked. This is the built-in Guardrails control aimed at unsupported answers in retrieval-based applications. The other options address jailbreak attempts, custom subjects, and exact strings, none of which measure whether an answer is supported.
Why the other options are wrong
- A. The prompt attack filter inspects user input for jailbreak or injection attempts. It does not evaluate whether the model's answer matches retrieved excerpts.
- C. Denied topics block subjects the team has defined. They do not score whether an answer is supported by source text.
- D. Word filters block literal strings from a list. An unsupported answer can contain none of those strings, so this control cannot detect it.