Databricks Generative AI Engineer Associate Question 144
Select 3You are designing a metaprompt for a generative AI model that summarizes sensitive legal documents. Your goal is to minimize hallucinations and prevent the leakage of private data. Which of the following strategies should you include in your metaprompt design?
- A
Instruct the model to only generate responses based on the provided context and avoid making assumptions.
- B
Include an explicit instruction for the model to hallucinate information when the context is unclear.
- C
Specify that the model should avoid using any external knowledge unrelated to the given input.
- D
Add a warning to the model to exclude all sensitive or private data from the output.
- E
Encourage the model to be creative and infer missing details when summarizing input.
Show answer and explanation
Correct answers: A, C, D
Explanation
To minimize hallucinations and avoid leaking private data, it is critical to design metaprompts that constrain the model's behavior. This includes grounding the model in the provided context, restricting the use of external knowledge, and explicitly instructing it to avoid generating private or sensitive information. Encouraging hallucination or creativity is counterproductive in sensitive tasks such as summarizing legal documents.
- A. Correct.
This is correct. Instructing the model to only use provided context reduces the likelihood of hallucinations by grounding its responses in the input data.
- B. Incorrect.
This is incorrect. Encouraging the model to hallucinate creates a high risk of generating inaccurate or fabricated information.
- C. Correct.
This is correct. Limiting the model to use only the given input and not external knowledge helps prevent hallucinations and keeps the response focused.
- D. Correct.
This is correct. Explicitly warning the model to exclude sensitive or private data ensures the output aligns with privacy requirements.
- E. Incorrect.
This is incorrect. Encouraging creativity and inference increases the chances of hallucinations and inaccurate outputs, which is undesirable in this context.