A compliance officer is reviewing the usage of Snowflake Cortex LLM functions and the Cortex REST API within their organization, specifically focusing on the implementation and impact of Cortex Guard. They observe several instances where 'guardrails' were enabled. Which of the following statements accurately describe the behavior and cost considerations of Cortex Guard when integrated with Snowflake Cortex LLM functions or the Cortex REST API?
Correct Answer: B,C,E
Option B is correct: when Cortex Guard is enabled and a response is blocked, the model's output is replaced by a message (defaulting to 'Response filtered by Cortex Guard'), and only 'guard_tokens' are counted as input tokens for Cortex Guard's processing, in addition to the primary LLM's prompt and completion tokens. Option C is correct as the argument allows customization of the filtered response message for both 'COMPLETE' and the REST API. Option E is correct because Cortex Guard is built with Meta's Llama Guard 3 and its usage is billed separately as 'guard_tokens' in addition to the 'COMPLETE function cost. Option A is incorrect because while guard_tokens' are billed, it's specifically for the guardrail processing, and the 'unsafe' completion tokens are not returned or billed as such, rather replaced by a filtered message. Option D is incorrect because Cortex Guard requires explicit enablement by setting 'guardrails' to 'TRUE'.