SnowPro Specialty: Gen AI Question 143
Single answerUse fully-managed LLMs, RAG, and text-to-SQL servicesA retail company stores sales, product, and inventory data in Snowflake and wants to let business analysts ask natural-language questions such as "Which product categories had the largest month-over-month revenue increase in EMEA last quarter?" The company also wants responses to be grounded in trusted business definitions, such as how "revenue" and "active product" are defined in internal policy documents. The team wants the fastest path using fully managed Snowflake capabilities, with minimal custom orchestration and no external vector database. Which solution best meets these requirements?
- A
Use Cortex Analyst to translate natural-language questions to SQL over the curated semantic model, and use Cortex Search over the policy documents to retrieve relevant business-definition context that can be included in the user experience.
- B
Fine-tune a custom open-source LLM outside Snowflake on the policy documents and structured tables, then call it from Snowflake with external functions for both SQL generation and document retrieval.
- C
Use Cortex Complete directly on raw table metadata and document text, without a semantic model or retrieval layer, because a larger prompt gives the model enough context to answer reliably.
- D
Export policy documents into a third-party vector store and build a custom RAG pipeline, because Snowflake managed services do not support retrieval over unstructured content together with text-to-SQL use cases.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to combine Cortex Analyst for text-to-SQL with Cortex Search for retrieval over internal policy documents. This reflects a practical Snowflake-native architecture for mixed structured and unstructured question answering. Cortex Analyst is intended for natural-language querying of structured enterprise data using a semantic model, which is important for expressing trusted business metrics, dimensions, relationships, and synonyms. Cortex Search provides managed retrieval over unstructured content, enabling RAG-style grounding from policy or glossary documents without requiring a separate vector database. Together, they support accurate analyst questions over tables while grounding responses in approved definitions. This approach is consistent with Snowflake best practices for using managed GenAI services: use a semantic layer for governed SQL generation, and use retrieval for document grounding rather than relying on a base LLM prompt alone.
- A. Correct.
Correct. Cortex Analyst is Snowflake's managed text-to-SQL capability designed to answer natural-language questions over structured data using a semantic model, which improves accuracy and governance for business metrics and joins. Cortex Search is Snowflake's managed retrieval service for unstructured content and can be used to ground answers or provide supporting policy-definition context from internal documents without introducing an external vector database. This combination aligns with the requirement for fully managed Snowflake services, minimal orchestration, and grounded responses.
- B. Incorrect.
Incorrect. This approach adds substantial operational complexity and moves key capabilities outside Snowflake, which conflicts with the requirement for the fastest path and minimal custom orchestration. It also assumes the need to fine-tune a custom model for tasks that Snowflake already addresses with managed services such as Cortex Analyst for text-to-SQL and Cortex Search for retrieval. A candidate might choose this option because custom models can appear more flexible, but it is not the best fit for the stated constraints.
- C. Incorrect.
Incorrect. Using a general LLM completion function alone over raw metadata and document text is not the recommended way to deliver reliable enterprise text-to-SQL. Without a semantic model, the model is more likely to misinterpret business definitions, join logic, and metric semantics. Without retrieval, long prompts become brittle and harder to maintain. This option reflects the common misconception that prompt size can replace structured semantic modeling and managed retrieval.
- D. Incorrect.
Incorrect. Snowflake does provide managed retrieval capabilities for unstructured content through Cortex Search, so an external vector store is not required for this scenario. Building a custom RAG pipeline may be appropriate in some architectures, but it does not best satisfy the stated requirement for fully managed Snowflake-native services with minimal custom orchestration. This distractor targets the misconception that Snowflake can only handle structured text-to-SQL and not enterprise retrieval.