A Generative AI Engineer has developed an LLM application to answer questions about internal company policies. The Generative AI Engineer must ensure that the application doesn't hallucinate or leak…

Question

Options

• AAdd guardrails to filter outputs from the LLM before it is shown to the user
• BFine-tune the model on your data, hoping it will learn what is appropriate and not
• CLimit the data available based on the user's access level
• DUse a strong system prompt to ensure the model aligns with your needs.

• A
  10% (3)
• B
  67% (20)
• C
  7% (2)
• D
  17% (5)

Explanation

Fine-tuning the model on internal policy data and 'hoping' it learns appropriate behavior is an unreliable and insufficient safeguard against both hallucination and confidential data leakage. Fine-tuning bakes knowledge into weights but provides no hard boundary-the model can still hallucinate or surface sensitive information in unexpected ways. It is not a control mechanism. By contrast, output guardrails (A) intercept and filter responses, access-level data restrictions (C) limit what the model can retrieve in the first place, and a strong system prompt (D) constrains the model's behavior at inference time. These are deterministic or near-deterministic controls; fine-tuning alone is not.

No community discussion yet for this question.