SnowPro Specialty: Gen AI Question 113
Single answerCortex AnalystA retail company is building a natural-language analytics assistant for regional sales managers. The team wants business users to ask questions such as "What were online sales in the Northeast last quarter?" and receive reliable SQL-based answers from Snowflake. They decide to use Cortex Analyst and need to minimize ambiguous metric definitions while keeping the solution maintainable as business terminology evolves. Which approach should the team take?
- A
Create and maintain a semantic model that defines business metrics, dimensions, relationships, and synonyms, and use Cortex Analyst to translate user questions into SQL against that model.
- B
Store a few example question-and-answer pairs in a table and rely on Cortex Analyst to infer metric definitions directly from raw transactional tables without additional modeling.
- C
Convert the transactional tables into vector embeddings and have Cortex Analyst retrieve similar rows to answer aggregate sales questions without using SQL.
- D
Use Cortex Search as the primary engine for numeric aggregation questions, because it is designed to calculate governed business metrics from unmodeled tables.
Show answer and explanation
Correct answer: A
Explanation
The best practice for Cortex Analyst is to ground natural-language analytical queries in a semantic model that captures business meaning explicitly. In this scenario, regional sales managers need reliable answers to questions involving governed metrics, time periods, channels, and geographies. A semantic model allows the team to define those concepts once and lets Cortex Analyst translate user questions into SQL consistently. This reduces ambiguity, improves trust, and makes the system easier to maintain as business vocabulary changes. In Snowflake documentation and product guidance, Cortex Analyst is positioned for conversational analytics over structured data using a semantic model, while Cortex Search serves semantic retrieval/search use cases rather than metric computation. The key applied takeaway is to use the right abstraction layer: semantic modeling for NL-to-SQL analytics, not embeddings or search alone.
- A. Correct.
Correct. Cortex Analyst is designed to answer natural-language analytical questions by generating SQL grounded in a semantic model. Defining metrics, dimensions, joins/relationships, and business synonyms in the semantic model helps reduce ambiguity and improves consistency for terms like "online sales," "Northeast," and "last quarter." This is also the most maintainable approach because terminology and definitions can be updated in the model rather than handled ad hoc in prompts or application code.
- B. Incorrect.
Incorrect. While examples can help in some AI workflows, relying on a handful of example Q&A pairs does not replace the governed semantic layer that Cortex Analyst uses to produce accurate, explainable SQL for analytics. Without explicit modeling of metrics and relationships, the system is more likely to misinterpret business terms or join logic. This option reflects the misconception that Analyst works best like a generic few-shot text model rather than a semantic-model-driven analytics capability.
- C. Incorrect.
Incorrect. Vector embeddings and similarity retrieval are useful for semantic search and retrieval scenarios, but they are not the right mechanism for governed analytical aggregation such as quarterly regional sales totals. Cortex Analyst answers these questions through SQL generation over structured data informed by a semantic model, not by retrieving semantically similar rows from embeddings.
- D. Incorrect.
Incorrect. Cortex Search is intended for low-latency semantic search and retrieval over text and similar content, not as the primary governed engine for business metric calculation from natural-language analytical questions. Numeric aggregations such as sales by region and time period are a Cortex Analyst use case when backed by a semantic model. This option confuses search/retrieval functionality with NL-to-SQL analytics.