SnowPro Specialty: Gen AI Question 160
Single answerCustom_instructions fieldA retail company is building a customer-support copilot in Snowflake and wants all responses generated through Cortex Analyst to follow company-specific rules: answer in a concise, professional tone, use the term "guest" instead of "customer," and avoid recommending refunds unless the semantic model explicitly contains refund-related policy data. The team wants these behavioral rules applied consistently without changing the user’s question each time. Which approach is the most appropriate?
- A
Populate the custom_instructions field with the company-specific response rules so the model applies them alongside the semantic model when generating answers.
- B
Add the company rules as extra dimensions and measures in the semantic model so Cortex Analyst treats them as mandatory generation constraints.
- C
Store the company rules only in the warehouse session parameters, because Cortex Analyst automatically reads session-level instructions before generating a response.
- D
Append the company rules manually to every end-user prompt, because Cortex Analyst does not support persistent behavioral guidance in the request payload.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use the custom_instructions field to provide persistent, request-level guidance for how Cortex Analyst should frame and constrain its responses. This is useful for enforcing organizational tone, vocabulary preferences, and answer-shaping rules that are not naturally part of the user’s question and do not belong in the semantic model itself. Best practice is to keep semantic models focused on data meaning and business logic, while using custom instructions for response behavior and presentation guidance. This separation helps maintain a clean model design and more predictable generated outputs. Refer to Snowflake Cortex Analyst documentation on request payload structure and the purpose of custom instructions for response guidance.
- A. Correct.
Correct. The custom_instructions field is intended for providing additional guidance on how the model should behave or present answers, beyond the user’s immediate question. In this scenario, tone, terminology preferences, and policy-oriented guardrails are exactly the kind of instructions that should be supplied through custom_instructions so they can be applied consistently across requests without forcing each user to rewrite the same requirements.
- B. Incorrect.
Incorrect. A semantic model defines business entities, metrics, relationships, and the data context used for analytical question answering. It is not the right mechanism for encoding stylistic or behavioral response rules such as tone, preferred wording, or caution around unsupported refund advice. Treating instruction text as dimensions or measures would be a misuse of the semantic model.
- C. Incorrect.
Incorrect. Session parameters can control Snowflake environment behavior, but Cortex Analyst does not automatically interpret warehouse session parameters as natural-language response instructions for generated answers. This option reflects a common misconception that all runtime context in Snowflake is automatically consumed by AI features.
- D. Incorrect.
Incorrect. Manually appending the same rules to every prompt could work operationally, but it is not the most appropriate approach here because the requirement is to apply consistent instructions without modifying every user question. The custom_instructions field exists specifically to avoid this repetitive and error-prone pattern.