SnowPro Specialty: Gen AI Question 74
Single answerCortex Agents (Public Preview)A retail company is building an internal support assistant using Snowflake Cortex Agents (Public Preview). The assistant must answer employee questions using product manuals stored in a Snowflake stage and current order-status data stored in Snowflake tables. During testing, the team notices that the agent answers procedural questions from the manuals well, but gives outdated or incomplete responses when users ask, "Where is order 12345 right now?" The architect wants the agent to ground responses in both unstructured files and structured business data while keeping the design aligned with Cortex Agents capabilities. Which approach is the BEST fit?
- A
Configure the Cortex Agent with both a Cortex Search service over the staged manuals and SQL access to the relevant order tables so the agent can retrieve document context and query live structured data
- B
Convert the order-status tables into PDF files, place them in a stage, and rely only on document retrieval so the agent has a single knowledge source
- C
Fine-tune the underlying LLM on historical order data every hour so the model memorizes current order status without querying Snowflake tables
- D
Create a larger Cortex Search index over the manuals only, because Cortex Agents cannot use structured Snowflake data during response generation
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use Cortex Agents as an orchestration layer that can combine retrieval from unstructured knowledge sources with access to structured Snowflake data. For product manuals in stages, a Cortex Search service is an appropriate way to provide relevant document context. For current order status, the agent should use live structured data rather than relying on static documents or model memory. This aligns with Snowflake guidance for agentic AI patterns: use retrieval-augmented generation for unstructured knowledge, and use tools or data access for up-to-date, queryable business facts. A common design mistake is trying to force structured operational data into documents or expecting an LLM to memorize changing facts. In practice, the strongest architecture grounds each question against the source that best matches the data type and freshness requirement.
- A. Correct.
Correct. Cortex Agents are designed to orchestrate retrieval and tool use across different data sources, including unstructured content and structured Snowflake data. In this scenario, the manuals are best handled through a retrieval layer such as Cortex Search, while order-status questions require querying live tables so responses reflect current transactional state. This is the most appropriate design for combining RAG over documents with access to fresh structured data.
- B. Incorrect.
Incorrect. Converting relational order data into PDFs introduces latency, duplication, and loss of the benefits of structured querying. Order status is dynamic and should typically be retrieved from live tables rather than periodically exported into documents. This option reflects a common misconception that all agent grounding must be converted into unstructured text.
- C. Incorrect.
Incorrect. Fine-tuning is not the right solution for rapidly changing operational data such as order status. Even if model customization were available in another context, retraining or fine-tuning on frequently changing records is inefficient and would still risk stale answers. Best practice is to keep volatile business facts in queryable data stores and let the agent retrieve them at runtime.
- D. Incorrect.
Incorrect. Making the document index larger does not solve the core problem of answering questions that depend on real-time structured records. The statement that Cortex Agents cannot use structured Snowflake data is the key error here. The scenario specifically calls for combining document retrieval with access to current structured data.