SnowPro Specialty: Gen AI Question 285
Single answerImplementation methodsA retail company is building a customer-support assistant in Snowflake. The team wants to answer questions using the company’s internal policy documents stored in Snowflake stages, while minimizing model hallucinations and avoiding the operational overhead of maintaining custom ML infrastructure. Which implementation method is the MOST appropriate?
- A
Build a retrieval-augmented generation (RAG) pipeline in Snowflake by chunking and embedding the internal documents, retrieving relevant chunks at query time, and passing the retrieved context to an LLM for answer generation.
- B
Fine-tune a large language model on the full set of policy documents so the model memorizes the content and no retrieval step is needed during inference.
- C
Export the staged documents to an external application, create a keyword index outside Snowflake, and call a general-purpose LLM directly without passing document excerpts in the prompt.
- D
Use a standard text-completion prompt with no document retrieval because hosted LLMs in Snowflake are already trained to answer enterprise-specific policy questions accurately.
Show answer and explanation
Correct answer: A
Explanation
The best implementation method is retrieval-augmented generation (RAG). In Snowflake Gen AI solutions, RAG is the practical choice when you need answers grounded in organization-specific data without the cost and lifecycle management of training or fine-tuning a model for every content update. The core pattern is: ingest documents, chunk them, create embeddings, retrieve the most relevant chunks for a user query, and send both the question and retrieved context to an LLM. This improves factuality and traceability compared with prompting alone. Fine-tuning is better suited to changing model behavior, formatting, or task specialization, not as the primary mechanism for keeping up with dynamic enterprise knowledge. This aligns with common best practices in Snowflake’s Cortex and document-grounding guidance, where retrieval-based approaches are recommended for question answering over proprietary data.
- A. Correct.
Correct. For enterprise knowledge-grounded question answering, RAG is typically the best implementation method when the goal is to use proprietary documents, reduce hallucinations, and avoid the cost and complexity of model retraining. In Snowflake, a practical pattern is to store documents in stages or tables, chunk them, generate embeddings, retrieve the most relevant chunks at inference time, and include that context in the prompt sent to an LLM. This approach keeps responses grounded in current enterprise content and is easier to update than retraining a model whenever documents change.
- B. Incorrect.
Incorrect. Fine-tuning is not the best fit for this scenario. Although fine-tuning can adapt model behavior or style for certain tasks, it is generally not the preferred implementation method for frequently changing knowledge bases such as policy documents. Fine-tuning also adds operational complexity and does not guarantee accurate recall of specific source content. A common misconception is that fine-tuning is the default solution for proprietary knowledge, but for document-based Q&A, retrieval-based grounding is usually more effective and maintainable.
- C. Incorrect.
Incorrect. Exporting data to an external system increases operational overhead, which the scenario explicitly wants to avoid. In addition, calling a general-purpose LLM without providing the retrieved document excerpts means the model is not grounded in the company’s actual policies, increasing the risk of hallucinated or outdated answers. Keyword-only indexing can also be less semantically effective than embedding-based retrieval for natural-language questions.
- D. Incorrect.
Incorrect. General-purpose hosted LLMs are not inherently aware of a company’s internal documents or latest enterprise policies unless that information is provided in context. Using prompting alone without retrieval is a common mistake for enterprise assistants. It may work for public knowledge tasks, but it is not appropriate when answers must be based on proprietary content stored in Snowflake.