SnowPro Specialty: Gen AI Question 118
Single answerCortex Agents (Public Preview)A retail company is piloting Snowflake Cortex Agents (Public Preview) to help support analysts answer questions such as, "Why did online returns spike in the Northeast last week?" The analysts want the agent to combine unstructured policy documents with structured sales and returns tables, and they also want the agent to decide when to use SQL versus document retrieval during a conversation. Which design best fits Cortex Agents for this use case?
- A
Create a Cortex Agent that is configured with access to both relevant semantic model/table resources and searchable document knowledge, allowing the agent to choose the appropriate tool during orchestration.
- B
Use a standard text-completion LLM prompt only, because Cortex Agents cannot orchestrate across both structured and unstructured enterprise data sources.
- C
Build a vector search application only, because Cortex Agents are limited to document retrieval and cannot answer questions that require SQL over tables.
- D
Train a custom foundation model inside Snowflake first, because Cortex Agents require model fine-tuning before they can reason over enterprise data.
Show answer and explanation
Correct answer: A
Explanation
The best answer is Option 1 because Cortex Agents are intended for agentic orchestration over enterprise data sources, including both structured and unstructured information. In a realistic analytics workflow, a question like "Why did online returns spike in the Northeast last week?" may require the agent to compute trends from tables and then ground the explanation in return-policy or operations documents. That is exactly the kind of mixed-data reasoning Cortex Agents are meant to support. By contrast, a prompt-only approach lacks reliable tool orchestration, a retrieval-only design cannot answer table-based analytical questions, and custom model training is not a prerequisite for Cortex Agents. This aligns with Snowflake guidance that Cortex Agents are used to connect model reasoning with enterprise data tools rather than relying solely on raw prompting.
- A. Correct.
Correct. Cortex Agents are designed to orchestrate across multiple data-access tools, including structured data access and unstructured retrieval, so they are a good fit when users ask mixed questions that may require SQL, retrieval, or both. In this scenario, the analysts need an agent that can reason over returns tables while also consulting policy documents, which aligns with the purpose of Cortex Agents in Public Preview.
- B. Incorrect.
Incorrect. This reflects a common misconception that a general LLM prompt alone is sufficient for enterprise analytics. A plain text-completion prompt does not provide the built-in orchestration and tool-use pattern needed to decide when to query structured data versus retrieve documents. Cortex Agents are specifically intended to coordinate those actions.
- C. Incorrect.
Incorrect. Vector retrieval alone can help with policy documents or other unstructured content, but it does not replace structured querying for metrics like regional return spikes over a defined time window. The scenario explicitly requires combining tables and documents, which is broader than a retrieval-only design.
- D. Incorrect.
Incorrect. Cortex Agents do not require customers to train or fine-tune a custom foundation model before use. The key task is configuring the agent with the right enterprise data resources and access patterns. Fine-tuning is not the prerequisite for enabling the agent to reason over enterprise data in this scenario.