SnowPro Specialty: Gen AI Question 48
Single answerCortex LLM functions (for example, task-specific, general)A support operations team stores incoming case data in a Snowflake table with columns CASE_ID, CUSTOMER_MESSAGE, and PRODUCT_LINE. They need to classify each message into one of five predefined issue categories for downstream routing, and they want the most reliable and maintainable approach using Snowflake Cortex LLM functions directly in SQL. Which approach should they choose?
- A
Use SNOWFLAKE.CORTEX.CLASSIFY_TEXT with the five issue categories as labels, because this task-specific function is designed for controlled text classification into predefined classes.
- B
Use SNOWFLAKE.CORTEX.COMPLETE and ask the model in a prompt to return exactly one of the five categories, because general-purpose generation is always more accurate than task-specific functions.
- C
Use SNOWFLAKE.CORTEX.SUMMARIZE on each customer message first, then map each summary to a category using a CASE statement, because summarization is the recommended first step for all classification pipelines.
- D
Use SNOWFLAKE.CORTEX.TRANSLATE to normalize all messages into English, and this alone will assign the most likely issue category without additional logic.
Show answer and explanation
Correct answer: A
Explanation
When selecting among Cortex LLM functions, a key design principle is to use task-specific functions for task-specific problems and reserve general functions for open-ended generation or cases where no specialized function fits. In this scenario, the team has a fixed set of known issue categories and needs deterministic routing-oriented classification, so SNOWFLAKE.CORTEX.CLASSIFY_TEXT is the most appropriate choice. Using SNOWFLAKE.CORTEX.COMPLETE for classification is possible but usually less maintainable because it relies on prompt design and output control rather than a function purpose-built for labeling. Similarly, SUMMARIZE and TRANSLATE solve different problems and do not replace classification. This aligns with Snowflake Cortex best practices: choose the function that most directly matches the intended NLP task to reduce complexity and improve operational reliability.
- A. Correct.
Correct. For a use case where the business already knows the allowed categories and wants consistent assignment into one of those categories, a task-specific Cortex function for classification is the best fit. CLASSIFY_TEXT is intended for labeling text against predefined classes, which typically improves maintainability and reduces prompt-engineering overhead compared with using a general text-generation function for the same purpose.
- B. Incorrect.
Incorrect. SNOWFLAKE.CORTEX.COMPLETE is a general-purpose text generation function and can sometimes be used for classification via prompting, but it is not the best choice when a task-specific function exists for controlled label assignment. The statement that general-purpose generation is always more accurate is a misconception; best practice is to prefer task-specific functions when they match the problem.
- C. Incorrect.
Incorrect. Summarization can be useful in some workflows, but it is not a recommended universal first step for classification. Adding SUMMARIZE introduces an extra transformation that can remove details needed for accurate categorization. For direct assignment into known labels, a classification function is more appropriate and simpler.
- D. Incorrect.
Incorrect. TRANSLATE is for language translation, not issue-category prediction. Translating content may be helpful as a preprocessing step in some multilingual scenarios, but it does not itself perform classification. Additional classification logic or a classification function would still be required.