SnowPro Specialty: Gen AI Question 117
Single answerCortex Agents (Public Preview)A retail company is building an internal support assistant with Cortex Agents (Public Preview). The assistant must answer employee questions by combining information from product manuals stored in a Snowflake table and current order status from a structured orders table. The team wants the agent to reason over both sources within a single response while keeping the implementation inside Snowflake as much as possible. Which approach best fits Cortex Agents for this requirement?
- A
Create a Cortex Agent that uses retrieval over the product-manual knowledge source and also invokes SQL tools against the orders table so the agent can combine unstructured and structured context in one answer.
- B
Use only AI_COMPLETE with a larger model prompt, because Cortex Agents cannot access either Snowflake tables or retrieved document context during generation.
- C
Export both tables to an external vector database and build the orchestration outside Snowflake, because Cortex Agents support only unstructured search and cannot incorporate SQL-based tools.
- D
Train a custom fine-tuned model on the product manuals and order history, because Cortex Agents require model retraining whenever answers must reference both documents and live table data.
Show answer and explanation
Correct answer: A
Explanation
The best answer is Option 1 because Cortex Agents (Public Preview) are designed for multi-step reasoning with tools and enterprise data inside Snowflake. In a realistic support-assistant scenario, the agent can use retrieval over unstructured product manuals and also query structured Snowflake tables for live order status, then synthesize both into one grounded response. This is preferable to relying only on a single completion call, exporting data for external orchestration, or trying to solve a live-data problem with fine-tuning. Snowflake guidance for Cortex Agents emphasizes agentic workflows that combine retrieval, tool use, and Snowflake-native data access patterns, which is exactly what this scenario requires.
- A. Correct.
Correct. Cortex Agents are intended for agentic workflows in Snowflake where an LLM can reason across multiple tools and data sources. A practical pattern is to give the agent access to retrieved knowledge from unstructured content, such as manuals, and structured access through SQL-oriented tools for live Snowflake tables like order status. This matches the scenario's need to blend document grounding with current transactional data in a single response while keeping orchestration in Snowflake.
- B. Incorrect.
Incorrect. AI_COMPLETE is useful for direct model inference, but by itself it does not provide the higher-level agent orchestration needed to select tools, retrieve supporting context, and combine structured and unstructured sources in a managed agent workflow. The statement that Cortex Agents cannot access Snowflake tables or retrieved context is the misconception here.
- C. Incorrect.
Incorrect. Although external architectures are possible in some solutions, this option conflicts with the requirement to keep the implementation inside Snowflake as much as possible. It also incorrectly assumes Cortex Agents can only do unstructured search. The key misunderstanding is that Cortex Agents are designed to coordinate more than one capability, including structured data access patterns.
- D. Incorrect.
Incorrect. Fine-tuning is not the right default solution for combining product documentation with live order data. The scenario requires access to current order status, which changes over time. Agentic retrieval and tool use are better suited than retraining a model whenever source data changes. This option reflects the common misconception that model training is required whenever multiple data sources must be used together.