ARA-C01 Question 197
Single answerCortex LLM functionsA financial services company stores customer support cases in Snowflake and wants to build an in-database workflow that classifies each case into one of several categories and produces a short summary for agents. The architects must minimize data movement, keep processing inside Snowflake, and make the solution easy to invoke from SQL pipelines. Which approach is the MOST appropriate?
- A
Use Snowflake Cortex LLM SQL functions such as COMPLETE to generate summaries and classify cases with prompt-based SQL queries directly against the case data in Snowflake.
- B
Export the support cases from Snowflake to an external application, call a third-party LLM API for both tasks, and write the results back because Cortex functions cannot be invoked from SQL.
- C
Create a materialized view that automatically calls an LLM for each new row, because Cortex LLM functions are designed to execute only inside materialized views for continuous inference.
- D
Use only standard SQL string functions and regular expressions, because Cortex LLM functions are limited to translation and cannot perform summarization or classification tasks.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use Snowflake Cortex LLM functions directly in SQL so the workflow remains inside Snowflake and can be incorporated into existing ELT and orchestration patterns. For this scenario, architects can use prompt-driven inference with Cortex functions to summarize case text and classify cases into business categories without exporting sensitive support data to another platform. This matches Snowflake best practices around reducing unnecessary data movement, simplifying architecture, and leveraging in-platform AI capabilities. Snowflake documentation for Cortex AI functions describes SQL-based invocation patterns for LLM-powered tasks such as text generation and summarization, which makes them suitable for production pipelines that need to operate close to the data.
- A. Correct.
Correct. Snowflake Cortex provides SQL-accessible LLM functions that can be used directly within Snowflake to perform generative AI tasks such as summarization and classification through prompt-based interactions. This approach aligns with the requirement to minimize data movement, keep processing in Snowflake, and integrate easily into SQL-based pipelines.
- B. Incorrect.
Incorrect. This approach increases data movement and operational complexity, which directly conflicts with the scenario requirements. It is also based on a false premise: Cortex LLM functions are available from SQL, so exporting data to an external application is not required for these use cases.
- C. Incorrect.
Incorrect. Materialized views are not the required or exclusive mechanism for invoking Cortex LLM functions. This option reflects a common misconception that LLM inference in Snowflake must be embedded into automatic view refresh logic. In practice, architects typically invoke Cortex functions from SQL statements, tasks, procedures, or application logic depending on the workflow.
- D. Incorrect.
Incorrect. Standard SQL string functions and regex are useful for deterministic text parsing, but they are not a substitute for LLM-based summarization and semantic classification. The statement that Cortex LLM functions are limited to translation is factually wrong; Snowflake Cortex supports broader generative tasks including text generation and summarization-oriented workflows.