Databricks Generative AI Engineer Associate Question 139
Select 3You are designing a metaprompt for a generative AI model to summarize documents while ensuring it does not hallucinate or inadvertently include sensitive private data. Which metaprompt strategy would best minimize these risks?
- A
Explicitly instruct the model to verify facts against a provided knowledge base before generating the summary.
- B
Ask the model to avoid including any information that is not directly found in the input document.
- C
Encourage the model to creatively infer missing details to produce a more engaging summary.
- D
Include a reminder in the prompt to exclude any personally identifiable information (PII) from the generated output.
- E
Avoid setting any constraints to improve the model's fluency and coherence.
Show answer and explanation
Correct answers: A, B, D
Explanation
To minimize hallucinations and prevent private data leakage, metaprompts should focus on grounding the model’s outputs in verifiable data, restricting the model to the input context, and explicitly instructing it to exclude sensitive information. Strategies that encourage creative inference or lack constraints can lead to inaccuracies and unintended data exposure.
- A. Correct.
This is correct because instructing the model to verify facts against a provided knowledge base can reduce hallucinations by grounding the output in factual data.
- B. Correct.
This is correct because explicitly stating that the model should not include information outside the input document minimizes the risk of hallucinating irrelevant or non-existent details.
- C. Incorrect.
This is incorrect because encouraging the model to infer missing details increases the risk of hallucinations and reduces control over the accuracy of the output.
- D. Correct.
This is correct because explicitly instructing the model to exclude PII addresses the risk of leaking private data.
- E. Incorrect.
This is incorrect because not setting constraints removes safeguards against hallucinations and private data leakage, increasing the likelihood of problematic outputs.