SnowPro Specialty: Gen AI Question 238
Single answer3.2 Set guardrails to filter out harmful or unsafe LLM responses.A healthcare analytics team is building a Snowflake-hosted assistant that uses Cortex LLM functions to answer internal questions from clinicians. During testing, the team finds that some prompts can produce unsafe output, such as dosage suggestions beyond approved policy or responses that include abusive language copied from user input. The team wants to reduce the risk of harmful responses while keeping the application architecture inside Snowflake as much as possible. Which approach is the MOST appropriate?
- A
Implement a review step that evaluates both user prompts and model outputs against safety criteria, and block, redact, or replace responses that violate policy before returning them to the user.
- B
Rely on a larger LLM model because higher-capability models do not generate unsafe responses when the underlying data is curated.
- C
Prevent unsafe responses by increasing the warehouse size so the model has more compute available to reason safely about regulated medical content.
- D
Remove all retrieval context from the application so the LLM cannot reference any potentially sensitive material, which eliminates the need for output filtering.
Show answer and explanation
Correct answer: A
Explanation
The strongest answer is to add explicit guardrails that inspect both inbound prompts and outbound completions, then enforce policy-based handling such as blocking, redacting, or returning a fallback response. For Snowflake GenAI solutions, best practice is not to depend solely on model selection or prompt wording for safety. Instead, teams should combine LLM functionality with application-level controls, moderation checks, and policy enforcement. This is especially important in regulated use cases like healthcare, where unsafe outputs can create material risk. In Snowflake documentation and broader GenAI best practices, safety patterns focus on layered controls: validate the input, constrain the task, evaluate the output, and log or monitor violations for governance and improvement.
- A. Correct.
Correct. The best practice for guardrails is to evaluate inputs and outputs against explicit safety policies and then enforce an action such as blocking, masking, or substituting a safe response. In a Snowflake-based GenAI application, this aligns with the practical pattern of adding moderation or policy checks around LLM calls rather than assuming the model alone will behave safely. Reviewing both prompts and completions is important because harmful content can originate from either user instructions or generated output.
- B. Incorrect.
Incorrect. A more capable or larger model may reduce some failure modes, but it does not eliminate unsafe, toxic, or policy-violating output. This option reflects a common misconception that model quality alone is a sufficient safety control. Guardrails should be explicit and enforceable, especially in regulated scenarios such as healthcare.
- C. Incorrect.
Incorrect. Warehouse size affects compute resources for Snowflake processing workloads, not whether an LLM response satisfies safety requirements. Increasing compute does not function as a guardrail and does not filter harmful or unsafe content. This distractor targets confusion between performance tuning and AI safety controls.
- D. Incorrect.
Incorrect. Removing retrieval context may reduce exposure to certain sensitive source material, but it does not address unsafe generations such as abusive language, fabricated medical guidance, or policy violations. Guardrails are still needed to screen prompts and outputs. This option confuses retrieval minimization with response safety enforcement.