SnowPro Specialty: Gen AI Question 101
Single answerSENTIMENTA retail company stores customer support chat transcripts in a Snowflake table and wants to monitor customer satisfaction trends by scoring each message as positive, negative, or neutral. The analytics team needs a solution that can be called directly in SQL, works on unstructured text without building a custom model, and can be used in scheduled transformations for dashboards. Which approach best meets these requirements?
- A
Use Snowflake Cortex SENTIMENT on the chat text in a SQL query and persist the returned label for downstream reporting.
- B
Train a custom sentiment classifier in Snowpark ML first, because Snowflake does not provide built-in sentiment analysis for SQL workloads.
- C
Use AI_COMPLETE with a prompt asking for sentiment, because SENTIMENT only works on structured numeric data.
- D
Convert each chat transcript into embeddings with EMBED_TEXT_768 and use the vector values directly as sentiment scores in the dashboard.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use Snowflake Cortex SENTIMENT directly in SQL. This is the most practical and maintainable approach when a team needs built-in sentiment analysis over unstructured text stored in Snowflake without creating and operating a separate custom model. In Snowflake best practices, purpose-built Cortex AISQL functions should be preferred when they directly address the task, because they simplify implementation and operationalization in SQL-centric pipelines. SENTIMENT is specifically intended to analyze text and return sentiment classification, making it appropriate for scheduled ELT jobs, reporting tables, and dashboard metrics. By contrast, using AI_COMPLETE for sentiment introduces unnecessary prompt dependency, embeddings do not directly provide sentiment outputs, and custom model training is only justified when built-in functionality does not meet business needs.
- A. Correct.
Correct. Snowflake Cortex provides a SENTIMENT function that can be invoked directly in SQL on text input to classify sentiment, making it suitable for operational SQL pipelines and scheduled transformations. This matches the requirement to avoid building a custom model and to integrate sentiment scoring into downstream reporting tables.
- B. Incorrect.
Incorrect. This is a common misconception. While Snowpark ML can be used for custom ML workflows, it is not required for basic sentiment analysis in Snowflake because Cortex includes a built-in SENTIMENT function for this exact use case. Training a custom model adds complexity and does not best satisfy the stated requirement for a direct SQL-based solution.
- C. Incorrect.
Incorrect. AI_COMPLETE can be prompted to perform many language tasks, but it is not the best fit here when a dedicated built-in sentiment function already exists. The claim that SENTIMENT only works on structured numeric data is false; SENTIMENT is designed for natural-language text.
- D. Incorrect.
Incorrect. Embeddings represent semantic meaning as vectors and are useful for similarity search, clustering, and retrieval scenarios, but the raw embedding values are not sentiment labels or sentiment scores. Additional modeling would be needed to derive sentiment from embeddings, so this does not directly meet the requirement.