SnowPro Specialty: Gen AI Question 234
Single answerIs data leaving/going to LLMs?A financial services company wants to let analysts summarize customer support cases using Snowflake Cortex, but its compliance team is concerned about whether sensitive case text leaves Snowflake and is sent to external LLM providers. The architect needs to choose an approach that best minimizes data egress to third-party model providers while still enabling LLM-powered summarization in Snowflake. Which option best addresses this concern?
- A
Use Snowflake Cortex functions with a model hosted by an external provider, because Snowflake guarantees that prompts never leave Snowflake when using any Cortex model.
- B
Use a model deployed in Snowflake as a Snowflake Native App, because all Native Apps include built-in LLM hosting and prevent data from leaving the Snowflake boundary.
- C
Use Snowflake Cortex functions and select a model that is hosted directly by Snowflake, because prompts and responses can stay within Snowflake-managed infrastructure rather than being routed to an external model provider.
- D
Use external functions instead of Cortex, because external functions automatically anonymize prompts before sending them to the LLM provider.
Show answer and explanation
Correct answer: C
Explanation
This question tests whether the candidate can distinguish between Snowflake-managed model inference and scenarios where prompts are routed to third-party providers. In Snowflake Cortex, data handling depends on the model and service architecture. If the organization's primary concern is whether data leaves Snowflake to reach an LLM provider, the best choice is to use a model hosted directly by Snowflake when that meets functional requirements. By contrast, assuming all Cortex usage keeps data entirely inside Snowflake is incorrect, and external functions clearly send data to external services. Native Apps also do not, by themselves, guarantee in-platform model hosting. Candidates should align model selection with data governance requirements and review Snowflake Cortex documentation on model availability, hosting, and data flow considerations when designing compliant GenAI solutions.
- A. Incorrect.
Incorrect. This reflects a common misconception. Snowflake Cortex provides access to models, but whether data is sent to an external model provider depends on the specific model being used. Snowflake does not guarantee that prompts never leave Snowflake for every Cortex model. If the selected model is provided by a third party, prompts may be processed by that provider according to the service design and documentation.
- B. Incorrect.
Incorrect. Snowflake Native Apps are a framework for distributing applications in Snowflake, not a mechanism that automatically provides LLM hosting for any app. Installing or using a Native App does not inherently mean model inference is hosted entirely within Snowflake, nor does it guarantee that no data leaves Snowflake.
- C. Correct.
Correct. To minimize data egress to third-party LLM providers, the architect should choose a Snowflake-hosted Cortex model when available. In that setup, inference is handled within Snowflake-managed infrastructure, which addresses the compliance concern more directly than using a model served by an external provider. This is the most appropriate design choice when the key requirement is reducing or avoiding sending prompts to third-party model vendors.
- D. Incorrect.
Incorrect. External functions invoke services outside Snowflake, typically through a cloud provider integration layer such as API Gateway and a remote service. They do not automatically anonymize prompts, and in fact they explicitly send data out of Snowflake to the external service. This would generally increase, not reduce, data egress concerns.