SnowPro Specialty: Gen AI Question 129
Single answerHelper functionsA data engineering team is building a support-ticket assistant in Snowflake. Incoming ticket text is stored in a table, and the team wants to classify each ticket into one of four predefined categories before passing it to a larger workflow. They want a solution that minimizes prompt-writing effort, enforces a constrained response format, and can be used directly in SQL transformations. Which approach is the most appropriate?
- A
Use a Cortex helper function designed for classification so the SQL query can return one of the predefined labels directly.
- B
Use COMPLETE with a long prompt that asks the model to classify the ticket, because helper functions cannot be called from SQL.
- C
Create a Python stored procedure that calls an external LLM API, because Snowflake helper functions are only for text generation and not structured tasks.
- D
Use EMBED_TEXT_768 on the ticket body and treat the resulting vector as the final category output without any additional logic.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use a Cortex helper function built for classification. Helper functions are intended to simplify common GenAI tasks in SQL by reducing prompt-engineering overhead and producing more structured, task-specific outputs than a general-purpose completion call. In a support-ticket pipeline, classification is a strong use case because the desired output is constrained to a known set of categories. While COMPLETE is more flexible, it is less ideal when the requirement is predictable labeling with minimal prompt-writing effort. Embedding functions serve a different purpose: they convert text into vectors for semantic comparison, not final labels. Snowflake documentation for Cortex AI SQL functions distinguishes between general-purpose generation functions and helper functions for targeted tasks, and best practice is to use the most specific function that matches the business requirement.
- A. Correct.
Correct. Snowflake Cortex provides helper functions for common GenAI tasks such as classification. In this scenario, a helper function is the best fit because it reduces the need for complex prompt engineering, can operate directly in SQL, and is designed to return constrained outputs aligned to a predefined set of labels. That makes it practical for production data pipelines where predictable outputs matter.
- B. Incorrect.
Incorrect. COMPLETE can sometimes be used for classification through careful prompting, but this approach requires more prompt design and output validation. The statement that helper functions cannot be called from SQL is false; helper functions are intended for use directly in SQL workflows. A candidate might choose this option because COMPLETE is flexible, but it is not the most appropriate option when a purpose-built helper function exists.
- C. Incorrect.
Incorrect. This introduces unnecessary architectural complexity and external dependencies. Snowflake helper functions are not limited to open-ended text generation; they are intended to simplify common tasks such as sentiment analysis, classification, summarization, and similar patterns. Someone might pick this option if they assume external APIs are needed for all GenAI use cases, but that is a misconception in the context of Cortex helper functions.
- D. Incorrect.
Incorrect. EMBED_TEXT_768 generates vector embeddings, which are useful for semantic search, retrieval, clustering, or downstream ML logic. An embedding vector is not itself a category label. To derive a class from embeddings, the team would still need additional similarity or classification logic. This option reflects a common misunderstanding that embeddings directly replace task-specific outputs.