ARA-C01 Question 198
Single answerCortex LLM functionsA global support organization stores customer case notes in a Snowflake table and wants to generate concise case summaries directly in SQL for downstream analytics. The architect must ensure the design minimizes data movement, uses Snowflake-managed large language models, and can be incorporated into standard SELECT statements over relational data. Which approach best meets these requirements?
- A
Use the SNOWFLAKE.CORTEX.COMPLETE function in a SQL query to generate summaries from the case note text stored in the table.
- B
Export the case notes to an external LLM endpoint through an external function, because Cortex LLM functions cannot be called from SQL against table data.
- C
Create a materialized view that invokes a Python UDF wrapping a third-party open-source model hosted on a warehouse, because Cortex functions are limited to unstructured stages only.
- D
Use a task to copy case notes into a separate vector database first, because Cortex LLM functions require embeddings and vector search before text generation is possible.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use SNOWFLAKE.CORTEX.COMPLETE directly in SQL. Snowflake Cortex provides LLM functions that can be invoked from SQL to perform generative AI tasks such as summarization over data already stored in Snowflake. This supports an architectural goal of keeping data in-platform, reducing egress, and simplifying governance. In contrast, exporting data to external LLMs or introducing custom model hosting adds unnecessary complexity and weakens the benefit of Snowflake-managed AI capabilities. Similarly, vector search and embeddings are optional patterns for retrieval use cases, not required for straightforward summarization. This aligns with Snowflake documentation and best practices for Cortex LLM functions: use built-in Cortex SQL functions when the task can be handled directly within Snowflake, especially when minimizing data movement and operational overhead is a primary requirement.
- A. Correct.
Correct. SNOWFLAKE.CORTEX.COMPLETE is designed for invoking Snowflake Cortex LLM capabilities directly from SQL. It can be used in queries against relational table data, allowing summaries or other generated text to be produced without exporting data to another platform. This aligns with the requirement to minimize data movement and use Snowflake-managed models in standard SQL workflows.
- B. Incorrect.
Incorrect. Although external functions can call outside services, this does not best meet the stated requirement. Cortex LLM functions are specifically intended to be called from SQL and can operate on text stored in Snowflake tables. Choosing an external endpoint would increase data movement and operational complexity, which is contrary to the scenario.
- C. Incorrect.
Incorrect. This option reflects a misconception that Cortex is limited to staged unstructured content. Snowflake Cortex LLM functions can be used directly with SQL expressions over relational data. Also, using a Python UDF with a self-managed model would add maintenance overhead and would not satisfy the requirement to use Snowflake-managed LLMs.
- D. Incorrect.
Incorrect. Embeddings and vector search are useful for retrieval-augmented scenarios, semantic search, or similarity matching, but they are not a prerequisite for text generation. If the goal is simply to summarize text already stored in a table, COMPLETE can be applied directly without first moving data into a vector database.