SnowPro Specialty: Gen AI Question 73
Single answerCortex Agents (Public Preview)A retail company is piloting a customer-support assistant built with Snowflake Cortex Agents (Public Preview). The assistant must answer order-status questions using the company's structured order tables and also reference return-policy documents stored in Snowflake stages. During testing, the team notices that responses are sometimes missing the latest order information even though the SQL query logic is correct. They want the agent to reliably use both structured and unstructured enterprise data while minimizing orchestration code in the application layer. Which implementation approach is the BEST fit?
- A
Configure a Cortex Agent with access to the relevant semantic model or SQL-accessible structured data and the staged return-policy documents, so the agent can plan across both sources and decide when to query or retrieve context during response generation.
- B
Replace the Cortex Agent with a single prompt sent directly to a completion model, embedding a nightly exported CSV of orders and the full return-policy text into the prompt so the model has all needed context.
- C
Use only Cortex Search over the return-policy documents and rely on the model to infer current order status from the user's question, since Cortex Agents are intended only for unstructured retrieval scenarios.
- D
Precompute static order summaries once per day and store them in a document collection for retrieval, because Cortex Agents cannot combine enterprise structured data access with unstructured document grounding in one workflow.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use Cortex Agents to orchestrate across both structured and unstructured data sources. In Snowflake's GenAI architecture, agents are intended for scenarios where a system must decide how to gather and use context from enterprise data rather than relying on a single static prompt. For this retail support use case, current order status should come from authoritative structured data, while return policies should be grounded in the staged documents. This approach aligns with best practices for reducing hallucinations, improving freshness of responses, and minimizing custom orchestration code. By contrast, embedding exported data in prompts or converting everything into static documents creates staleness and scale problems. Similarly, using only unstructured retrieval cannot reliably answer transactional questions that require live structured data.
- A. Correct.
Correct. Cortex Agents (Public Preview) are designed to orchestrate across enterprise data sources and tools, including structured data access and unstructured document retrieval, to answer user questions with grounded context. In this scenario, the business needs up-to-date order data from structured tables plus policy information from documents. Using an agent is the best fit because it can decide when to query current structured data and when to retrieve policy content, reducing custom orchestration in the application.
- B. Incorrect.
Incorrect. Sending a single prompt with embedded CSV and policy text is brittle, quickly hits context limitations, and does not ensure fresh order data unless the export is constantly regenerated. It also shifts orchestration burden to the application rather than using the agentic capabilities intended for combining enterprise data sources.
- C. Incorrect.
Incorrect. Cortex Search is useful for retrieving relevant unstructured content, but it does not replace access to live structured order data. The misconception is assuming the model can infer current transactional facts without querying the authoritative tables. For order status, the system needs grounded, current structured data access.
- D. Incorrect.
Incorrect. Precomputing static summaries introduces staleness, which is exactly the issue the team is trying to solve. The statement that Cortex Agents cannot combine structured and unstructured enterprise data in one workflow is also false; that combination is a primary reason to use an agent in this scenario.