SnowPro Specialty: Gen AI Question 14
Single answerCortex Agents (Public Preview)A retail company is piloting Snowflake Cortex Agents (Public Preview) to support internal analysts who ask natural-language questions such as, "Which product lines saw declining margin after the spring campaign, and what inventory issues were mentioned in support tickets?" The team wants the agent to ground answers in both structured warehouse tables and unstructured ticket documents, while minimizing hallucinations and avoiding custom orchestration code outside Snowflake. Which design is the best fit for this requirement?
- A
Build a Cortex Agent that can use multiple Snowflake tools, such as SQL over structured data and search/retrieval over relevant documents, so the model can reason across both sources during answer generation.
- B
Fine-tune a base LLM on exported warehouse tables and support tickets, then deploy it outside Snowflake so it memorizes both structured and unstructured business context.
- C
Create a standard semantic view for the structured data and require analysts to manually paste relevant support ticket text into each prompt when they need document context.
- D
Use a single vector search index for all data, including numeric warehouse facts, and rely on semantic similarity alone instead of letting the agent invoke SQL.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use Cortex Agents as a tool-orchestration layer inside Snowflake so the model can access the right data source in the right way. For structured warehouse data, SQL remains the most reliable mechanism for exact analytics. For unstructured content like support tickets, search/retrieval provides relevant grounding. This combination is especially important in enterprise scenarios where users ask compound questions spanning metrics and text evidence. Snowflake guidance for Cortex Agents emphasizes tool use and grounded responses rather than trying to force a single retrieval method to handle every workload. In practice, using SQL for facts and retrieval for documents reduces hallucinations, improves freshness, and avoids external custom orchestration.
- A. Correct.
Correct. Cortex Agents are designed to orchestrate tools in Snowflake so an LLM can answer questions using grounded enterprise data. In this scenario, the key requirement is to combine structured analysis from warehouse tables with unstructured context from support tickets, while avoiding custom orchestration outside Snowflake. A Cortex Agent can route to appropriate tools such as SQL for precise aggregation/filtering on structured data and search/retrieval for document grounding. This is the most practical and accurate pattern for minimizing hallucinations and answering multi-source business questions.
- B. Incorrect.
Incorrect. Fine-tuning is not the best solution for this scenario. Exporting enterprise data and training a model to memorize it introduces governance, freshness, and maintenance issues. It also does not align with the requirement to avoid custom orchestration outside Snowflake. For business questions that depend on current data and documents, grounded tool use and retrieval are generally preferred over training a model to memorize changing facts.
- C. Incorrect.
Incorrect. A semantic layer or curated structured interface can help with governed SQL access, but this option fails the main requirement to combine structured and unstructured sources without manual effort. Requiring users to paste ticket text into prompts is operationally weak, error-prone, and does not scale. It also increases the risk of inconsistent grounding and user-dependent results.
- D. Incorrect.
Incorrect. Treating all data as vectorized text is a common misconception. Vector retrieval can help find relevant document context, but it is not a substitute for SQL when users need exact calculations, filters, joins, trends, or aggregations on structured warehouse data. Relying only on semantic similarity for numeric business facts can reduce accuracy and make answers less trustworthy.