SnowPro Specialty: Gen AI Question 6
Single answerLLMsA retail company is building a customer-support assistant in Snowflake that uses a large language model (LLM) to answer questions from product manuals, warranty documents, and return policies. During testing, the team notices that the model sometimes produces fluent but incorrect answers when the requested information is not present in the source documents. They want to reduce these hallucinations without fine-tuning the base model and while keeping the implementation maintainable. Which approach is the BEST fit for this requirement?
- A
Implement retrieval-augmented generation (RAG) so the application retrieves relevant document chunks at query time and includes them in the prompt, while instructing the model to answer only from the provided context
- B
Increase the temperature setting so the model explores a wider range of possible answers and is less likely to repeat incorrect patterns
- C
Replace the LLM with a larger parameter model and rely on its pretrained knowledge instead of grounding responses in company documents
- D
Fine-tune the model on a small sample of support tickets so it memorizes the correct answers for future prompts
Show answer and explanation
Correct answer: A
Explanation
The best answer is to implement retrieval-augmented generation (RAG). In enterprise LLM applications, hallucinations often occur because the model is asked to answer questions that require current, proprietary, or highly specific information that is not reliably embedded in the model's pretrained parameters. RAG addresses this by retrieving relevant data from a trusted knowledge base at runtime and passing it as context to the model. This pattern is widely considered a best practice for question answering over business documents because it improves factual grounding, supports changing source content without retraining, and keeps the system more maintainable than repeated fine-tuning cycles. In Snowflake-based GenAI architectures, this aligns with using data stored in Snowflake as the source of truth and building LLM workflows that retrieve and condition on enterprise data rather than relying solely on the base model's prior knowledge.
- A. Correct.
Correct. RAG is a standard pattern for reducing hallucinations in enterprise question-answering use cases. By retrieving the most relevant passages from trusted documents at inference time and supplying that context to the model, the application grounds the response in authoritative data. Adding prompt instructions such as answering only from supplied context and indicating when the answer is unavailable further improves reliability. This approach also avoids the operational cost and maintenance burden of model fine-tuning when the main issue is missing or changing factual context.
- B. Incorrect.
Incorrect. Higher temperature generally increases randomness and creativity, which is usually the opposite of what you want when trying to reduce hallucinations in factual support workflows. For grounded question answering, lower temperature is often preferred because it encourages more deterministic outputs. The misconception here is assuming that more variation helps correctness; in practice, it often makes factual consistency worse.
- C. Incorrect.
Incorrect. A larger model may improve overall reasoning or language quality, but it does not solve the core problem: the model is answering without reliable grounding in the company's current documents. Pretrained knowledge can be outdated, incomplete, or irrelevant to proprietary policies. This option reflects the common misconception that model size alone is the best way to improve factual accuracy in enterprise-specific tasks.
- D. Incorrect.
Incorrect. Fine-tuning can help shape style or task behavior in some cases, but it is not usually the best first solution when the issue is access to up-to-date source content. A small support-ticket dataset may not cover all policy and product-document variations, and fine-tuning does not inherently provide live access to changing documents. This also adds lifecycle complexity compared with a retrieval-based design.