SnowPro Specialty: Gen AI Question 258
Single answerCortex LLM functionsA retail analytics team stores customer reviews in a Snowflake table and wants to build a SQL-only pipeline that classifies each review into one of these labels: SHIPPING, PRODUCT_QUALITY, BILLING, or OTHER. The team needs the model to return one label per row in a predictable format so downstream SQL logic can aggregate counts by category without extensive post-processing. Which Cortex LLM function is the most appropriate choice for this requirement?
- A
Use SNOWFLAKE.CORTEX.COMPLETE with a prompt asking the model to output exactly one of the allowed labels
- B
Use SNOWFLAKE.CORTEX.SUMMARIZE to condense each review, then map the summary to a category in a second step
- C
Use SNOWFLAKE.CORTEX.CLASSIFY_TEXT with the provided category labels to assign each review to one of the target classes
- D
Use SNOWFLAKE.CORTEX.TRANSLATE to normalize review language before assigning categories directly from the translation output
Show answer and explanation
Correct answer: C
Explanation
This scenario is about selecting the most appropriate Cortex LLM function for a practical SQL workflow. When a team has predefined classes and needs deterministic, structured categorization for downstream analytics, a purpose-built classification function is the best choice. SNOWFLAKE.CORTEX.CLASSIFY_TEXT is designed for text classification tasks, whereas SNOWFLAKE.CORTEX.COMPLETE is a general LLM text generation function that can be adapted through prompting but is not the most direct or reliable option for constrained category assignment. SUMMARIZE and TRANSLATE address different NLP tasks and would introduce unnecessary steps. This reflects a common best practice in Snowflake Cortex: choose the specialized function that matches the business task rather than forcing a general completion function to emulate it.
- A. Incorrect.
Incorrect. COMPLETE can often be prompted to return a single label, but it is a general-purpose text generation function rather than the most appropriate purpose-built option for controlled text classification. In production, relying only on prompt wording can be less predictable than using a specialized classification function designed for assigning text into predefined classes.
- B. Incorrect.
Incorrect. SUMMARIZE is intended to produce condensed versions of text, not classify text into user-defined categories. While a team could add later logic to infer a category from a summary, that adds unnecessary complexity and reduces reliability compared to direct classification.
- C. Correct.
Correct. CLASSIFY_TEXT is the best fit when the task is to assign input text to one of a predefined set of labels. In this scenario, the team needs one category per review in a structured, downstream-friendly format. A dedicated classification function aligns directly to that requirement and is more appropriate than using a general completion model for label enforcement.
- D. Incorrect.
Incorrect. TRANSLATE is intended for language translation, not sentiment or topic classification. Translating text may help standardize multilingual inputs in some workflows, but translation output does not directly solve the requirement to map each review to one of the specified business categories.