SnowPro Specialty: Gen AI Question 87
Single answerGeneralA retail company is building an internal assistant in Snowflake to help support agents answer questions about return policies, shipping rules, and warranty terms. The source content exists as unstructured PDF and HTML documents stored in Snowflake stages, and the company wants responses to be grounded in its own documents rather than relying only on a general-purpose model's pretrained knowledge. Which approach should the team take to best meet this requirement?
- A
Use Cortex Search over the indexed document content and pass the retrieved relevant chunks to a large language model to generate the final answer
- B
Fine-tune a large language model directly on the staged PDF and HTML files so the model memorizes the policy documents
- C
Use a standard SQL view over file metadata and ask a large language model to answer questions without retrieving document content
- D
Rely on the base model alone because general-purpose models are designed to answer enterprise policy questions accurately without additional context
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use retrieval-augmented generation with Cortex Search and an LLM. In Snowflake GenAI patterns, enterprise assistants typically need grounded responses based on proprietary or frequently changing content. Retrieval addresses this by finding the most relevant chunks from the organization's documents at query time and supplying them to the model as context. This is generally preferable to relying solely on pretrained model knowledge, which may be incomplete or outdated, and preferable to fine-tuning when the primary need is accurate use of source documents rather than changing the model's style or task behavior. Snowflake documentation and best practices for AI applications emphasize grounding model outputs with relevant enterprise data for higher accuracy and trustworthiness.
- A. Correct.
Correct. This is the standard retrieval-augmented generation (RAG) pattern for enterprise question answering. Cortex Search is designed to retrieve relevant content from a knowledge base, and those retrieved passages can then be supplied to an LLM so responses are grounded in current company documents. This approach improves factuality, supports updates to source content without retraining the model, and is well suited for unstructured documents such as PDFs and HTML that have been prepared for search.
- B. Incorrect.
Incorrect. Fine-tuning is not the best first choice for this requirement, especially when the goal is to ground answers in changing enterprise documents. Fine-tuning changes model behavior but is not primarily a document retrieval mechanism, and it is not an efficient way to keep answers aligned with frequently updated policy content. A candidate might choose this option because training sounds like a way to 'teach' the model company knowledge, but in practice RAG is the more appropriate pattern for document-grounded Q&A.
- C. Incorrect.
Incorrect. A SQL view over file metadata does not provide the substantive document text needed to answer policy questions. Even if metadata such as filename, path, or timestamp is available, the model still needs relevant content passages from the actual documents. This distractor reflects a common misconception that cataloging files is enough for semantic question answering.
- D. Incorrect.
Incorrect. Base models can produce fluent answers, but without retrieval from enterprise documents they may omit company-specific nuances, use outdated information, or hallucinate details. For internal policy assistants, grounding with trusted source content is a best practice. This option represents the mistaken belief that pretrained knowledge alone is sufficient for domain-specific enterprise support use cases.