NCA-GENL Trustworthy AI Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✓

  Colang flows that specify dialogue patterns and guardrail actions.

  Why this is correct

  Colang is the modeling language used to define conversational flows and guardrails in NeMo Guardrails. By writing Colang flows, the team can specify allowed and disallowed topics, enforce grounding to the approved corpus, and trigger actions like blocking or redirecting responses. This directly addresses the requirement to constrain behavior and prevent unapproved content.

• ✗

  A custom embedding model fine-tuned on the document corpus.

  Why it's wrong here

  Fine-tuning an embedding model improves retrieval relevance but does not enforce behavioral constraints or block unapproved content. It cannot define dialogue patterns or guardrail actions. While it can help ground responses, it lacks the control logic to prevent jailbreaks or off-topic responses, which are central to the team's requirement.

• ✗

  TensorRT-LLM optimization profiles for inference acceleration.

  Why it's wrong here

  TensorRT-LLM optimizes inference speed and throughput but has no role in defining or enforcing content policies. It does not provide mechanisms to ground responses or block disallowed content. Using it would improve performance, not trustworthiness constraints, so it fails to meet the stated objective.

• ✗

  The NVIDIA Triton Inference Server model ensemble configuration.

  Why it's wrong here

  Triton ensembles orchestrate multiple models for inference but do not define conversational guardrails or content restrictions. They cannot specify allowed topics or block unapproved outputs. While useful for serving, they do not address the need for behavioral constraints and grounding to the approved corpus.