SnowPro Specialty: Gen AI Question 88
Single answerGeneralA retail company wants to build a customer-support assistant in Snowflake using Snowflake Cortex. The assistant must answer questions using the company's policy documents, return grounded responses, and minimize the risk of fabricated answers. The team also wants the solution to stay within Snowflake-managed capabilities as much as possible rather than building a custom orchestration layer outside Snowflake. Which approach should the architect recommend?
- A
Use a retrieval-augmented generation pattern in Snowflake by retrieving relevant document chunks from company data and passing that context to a Cortex LLM for answer generation.
- B
Fine-tune a Cortex model directly on all policy documents so the model permanently memorizes the content and no document retrieval is needed at runtime.
- C
Send every customer question directly to a general-purpose Cortex LLM with no grounding context because Snowflake-hosted models automatically avoid hallucinations on enterprise-specific content.
- D
Export policy documents to an external vector database and application stack because Snowflake does not support running grounded generative AI workflows with enterprise data.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use retrieval-augmented generation (RAG) with Snowflake-managed generative AI capabilities. For enterprise question-answering, best practice is to retrieve relevant internal content at runtime and provide it to the model as grounding context. This improves answer quality, keeps responses aligned to current source material, and reduces hallucinations compared with relying on a model's general pretraining alone. In Snowflake GenAI guidance, Cortex is positioned as a way to build AI applications over enterprise data, and grounded generation is a core practical pattern for document-based assistants. Fine-tuning is not the default solution for dynamic knowledge bases like policy documents, and ungrounded prompting is risky for private, organization-specific answers.
- A. Correct.
Correct. A retrieval-augmented generation (RAG) approach is the best fit for this scenario because it grounds the model's response in relevant enterprise documents at inference time. This reduces hallucination risk compared with prompting an LLM without context and aligns with the requirement to stay within Snowflake-managed capabilities. In practice, the workflow retrieves relevant chunks from the company's data and supplies them as context to a Cortex model for answer generation.
- B. Incorrect.
Incorrect. The idea that teams should fine-tune a Cortex model directly on all enterprise documents to avoid retrieval is a common misconception. Fine-tuning is not the standard recommendation for rapidly changing policy content, and it is not the primary Snowflake-managed pattern for grounded question answering over enterprise documents. Policy documents change over time, so retrieval at runtime is generally better for freshness and governance.
- C. Incorrect.
Incorrect. Even high-quality LLMs can hallucinate, especially when asked about organization-specific content not included in the prompt context. Snowflake-hosted models do not inherently know a company's private policy documents unless that information is provided through context or another supported grounding mechanism. Sending questions directly to the model without grounding would not meet the stated requirement.
- D. Incorrect.
Incorrect. This option is based on the false premise that Snowflake cannot support grounded generative AI workflows with enterprise data. Snowflake provides managed generative AI capabilities, including Cortex-powered functions and patterns for using enterprise data in AI applications. An external stack may be possible in some architectures, but it is not required here and conflicts with the goal of maximizing Snowflake-managed capabilities.