AIncorrect
Content filter for harmful language
A content filter is a feature that blocks harmful language such as hate and violence.
It controls harmfulness, not the verification of consistency with reference documents, so it is incorrect.
A company is strengthening hallucination defenses for a RAG chatbot. Which Bedrock guardrails mechanism verifies whether a RAG-generated response is actually grounded in the content of the provided reference documents (sources), that is, whether it strays from the evidence?
1 / 1
Select an answer
CorrectD
Explanation
Choosing the guardrails mechanism that verifies a response's grounding.
- 1「actually grounded in the content of the provided reference documents (sources)」Confirm whether it follows the evidence
- 2「Which Bedrock guardrails mechanism verifies」Evidence verification = contextual grounding check
BIncorrect
Denied topics
Denied topics is a feature that blocks topics you do not want handled by specifying them.
It restricts topics, not verifying whether a response is grounded in reference documents, so it is incorrect.
CIncorrect
Word filter
A word filter is a feature that blocks the appearance of specified words or phrases.
It controls at the word level, not verifying whether the whole response is grounded in evidence, so it is incorrect.
DCorrect
Contextual grounding check
Correct. The contextual grounding check is a Bedrock guardrails mechanism that verifies whether a response is grounded in the content of the provided reference documents (sources) and suppresses output that strays from the evidence.
Key Takeaway
Remember the correct answer, the contextual grounding check.
・A Bedrock guardrails mechanism that verifies whether a response is grounded in the content of the provided reference documents (sources).
・It detects and suppresses output that strays from the evidence (contradicts facts), improving the reliability of RAG.