SnowPro Specialty: Gen AI Question 196
Single answer2.5 Run third-party models in Snowflake.A data science team wants to use a third-party large language model from within Snowflake to summarize customer support cases. The security team requires that application code and SQL users call the model through a controlled interface, and the platform team wants to avoid moving data out of Snowflake pipelines manually. Which approach best meets these requirements?
- A
Create a Cortex Search service and use it directly to generate summaries from support case text.
- B
Create an external function that invokes the third-party model provider through a managed integration path, then expose that function to analysts and applications.
- C
Export support case data to an external stage, run the model outside Snowflake, and load the summaries back into Snowflake tables on a schedule.
- D
Create a dynamic table that calls the third-party model endpoint directly over the public internet without any integration object.
Show answer and explanation
Correct answer: B
Explanation
The best answer is to use an external function to call the third-party model through Snowflake's supported integration architecture. This is the practical pattern for running third-party services from SQL while maintaining governance, abstraction, and operational consistency. It allows developers to present a stable SQL interface to downstream users and applications, instead of forcing data exports or bespoke application-side orchestration. By contrast, Cortex Search is for retrieval use cases, not arbitrary third-party model execution. Exporting data out and re-importing results adds unnecessary movement and complexity. Direct unmanaged outbound internet calls from objects like dynamic tables are not the supported way to invoke external model providers. This aligns with Snowflake guidance on using external functions and integrations when connecting Snowflake workloads to external services.
- A. Incorrect.
Incorrect. Cortex Search is designed for retrieval and search over indexed content, not as a generic mechanism to run arbitrary third-party generative models for summarization. A candidate might choose this if they confuse retrieval features with model inference features.
- B. Correct.
Correct. Using an external function is the standard Snowflake pattern when you need SQL-accessible calls to an external service, including a third-party model endpoint, while keeping invocation behind a controlled interface. This approach fits the requirement to let SQL users and applications call the model without manually exporting data from Snowflake workflows. In practice, external functions are used with the appropriate integration components so calls are managed and governed rather than embedded as ad hoc outbound requests.
- C. Incorrect.
Incorrect. This would work functionally, but it does not best meet the stated requirements. It introduces manual or batch-oriented data movement outside Snowflake pipelines, increases operational complexity, and weakens the goal of providing a controlled in-platform interface for SQL users and applications.
- D. Incorrect.
Incorrect. Snowflake objects such as dynamic tables are not a mechanism for making arbitrary direct public internet API calls. Third-party service access from Snowflake requires supported integration patterns rather than embedding unmanaged outbound network access in SQL transformations.