CCAO-F Claude Model Fundamentals 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.

• ✓

  Add a system prompt that instructs Claude to only use the provided knowledge base and to say 'I don't know' when the answer is not found.

  Why this is correct

  A system prompt sets persistent, high-level behavioral instructions that Claude prioritizes throughout the conversation. By explicitly restricting Claude to the provided knowledge base and requiring an 'I don't know' response when information is absent, the system prompt reliably constrains the model's scope. This is the recommended approach for grounding and refusal behavior in Claude on Amazon Bedrock, as it leverages Claude's instruction-following strength without altering the model weights.

• ✗

  Increase the max_tokens parameter so Claude has enough room to include the full knowledge base in every response.

  Why it's wrong here

  max_tokens limits the length of the generated response, not the input context. Raising it allows longer outputs but does nothing to constrain Claude to the knowledge base or to trigger refusals. Including the entire knowledge base in the output is also impractical and unrelated to grounding. This parameter has no effect on whether Claude stays within the provided information.

• ✗

  Fine-tune the Claude 3.5 Sonnet model on the knowledge base so that it memorizes the content and refuses out-of-scope questions.

  Why it's wrong here

  Fine-tuning Claude 3.5 Sonnet on Amazon Bedrock is not supported for this use case, and even if it were, fine-tuning teaches style and format rather than reliable factual grounding. The knowledge base changes frequently, so retraining would be impractical. It also risks catastrophic forgetting and does not guarantee refusal behavior. A system prompt is the correct, supported mechanism for constraining responses to a knowledge source.

• ✗

  Lower the temperature setting to 0 so Claude becomes deterministic and will not hallucinate outside the knowledge base.

  Why it's wrong here

  Temperature controls randomness in token selection, not factual grounding. At temperature 0, Claude produces the most likely token sequence, but it can still generate plausible-sounding content that is not in the knowledge base. Determinism does not equal accuracy. The model has no inherent awareness of the knowledge base boundary, so temperature alone cannot enforce refusal behavior or restrict answers to provided content.