SnowPro Specialty: Gen AI Question 181
Single answerSnowflake CortexA retail company stores product reviews in a Snowflake table and wants to add a SQL-only pipeline that classifies each review by sentiment and generates a short summary for customer-service agents. The data team does not want to manage external model endpoints or move data out of Snowflake. Which approach best meets these requirements using Snowflake Cortex?
- A
Use Snowflake Cortex AISQL functions such as SENTIMENT and SUMMARIZE directly in SQL against the review text stored in Snowflake tables.
- B
Export the reviews to an external vector database, call a third-party hosted LLM through custom Python code, and write the results back to Snowflake.
- C
Create a Snowpark Container Services application because Cortex functions require containerized deployment before they can be called from SQL.
- D
Train a custom sentiment model inside a virtual warehouse, because Cortex only supports text generation and cannot perform classification-style tasks.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use Snowflake Cortex AISQL functions directly in SQL. This approach satisfies all stated constraints: it keeps data inside Snowflake, avoids managing external LLM infrastructure, and supports practical text-processing tasks like sentiment analysis and summarization. In Snowflake documentation, Cortex is positioned as a managed AI capability that lets users invoke LLM-powered and task-specific functions from SQL for workflows on data already stored in Snowflake. For this scenario, using functions such as SENTIMENT and SUMMARIZE is more operationally efficient and more aligned with Snowflake best practices than exporting data or building custom serving layers.
- A. Correct.
Correct. Snowflake Cortex provides AI functions that can be invoked directly from SQL, enabling teams to analyze unstructured text without building and managing external model-serving infrastructure. Functions such as SENTIMENT and SUMMARIZE are designed for exactly this kind of in-database workflow, which aligns with the requirement to keep data in Snowflake and avoid managing external endpoints.
- B. Incorrect.
Incorrect. While this could work technically, it violates key requirements in the scenario: the team does not want to manage external model endpoints or move data out of Snowflake. Exporting data to an external vector database and orchestrating third-party LLM calls adds operational overhead, security considerations, and latency that Cortex is intended to reduce for common AI use cases.
- C. Incorrect.
Incorrect. Snowpark Container Services is used for packaging and running custom applications or models, but it is not required to use Snowflake Cortex SQL functions. A common misconception is that all GenAI workloads in Snowflake need custom runtime infrastructure. In this scenario, built-in Cortex functions are sufficient and simpler.
- D. Incorrect.
Incorrect. Cortex is not limited to free-form text generation. It includes task-oriented AI capabilities for common text operations such as sentiment analysis, summarization, translation, and related inference use cases. The misunderstanding here is assuming that only custom model training can support classification-like outcomes, when Cortex already exposes managed functions for these tasks.