NCP-GENL Prompt Engineering 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.

• ✗

  Fine-tune the model on the internal policy documents, then use a simple zero-shot prompt to ask questions.

  Why it's wrong here

  Fine-tuning can adapt a model to a domain, but it is a heavyweight, offline process and does not guarantee that the model will answer only from the current policy text. Policies change frequently, and fine-tuning would need to be repeated. For an interactive assistant requiring up-to-date, grounded answers, in-context grounding is more appropriate and directly addresses the requirement.

• ✗

  Use a chain-of-thought prompt that asks the model to reason step by step about the policy before answering.

  Why it's wrong here

  Chain-of-thought can improve reasoning on complex tasks, but it does not ground the model in a specific set of internal policies. Without the actual policy text in the context, the model may still generate plausible but incorrect policy statements. The scenario requires the model to answer from a provided set of policies, so context injection with a strict instruction is the correct approach here.

• ✗

  Increase the temperature setting to 0.9 to encourage the model to explore a wider range of policy interpretations.

  Why it's wrong here

  Higher temperature increases randomness and creativity, which is the opposite of what is needed for factual policy Q&A. It would make the model more likely to invent or misstate policy details. For grounded, accurate answers, a low temperature (e.g., 0.1) is typically preferred, and temperature alone does not constrain the model to the provided policy context.

• ✓

  Provide the relevant policy excerpts directly within the prompt, instruct the model to answer only from that context, and to state when the answer is not present.

  Why this is correct

  This is retrieval-augmented generation (RAG) at the prompt level. By injecting the policy excerpts into the context and explicitly constraining the model to use only that information, you prevent it from relying on its pre-trained knowledge. Instructing it to say when the answer is missing further reduces hallucination and keeps the response grounded in the provided internal policies.