SnowPro Specialty: Gen AI Question 5
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 terms, and return-policy documents. During testing, the team finds that answers sometimes sound fluent but include policy details that are not present in the source documents. The team wants to reduce these hallucinations while still allowing the model to generate natural-language responses. Which approach is the MOST appropriate?
- A
Implement retrieval-augmented generation (RAG) so the application retrieves relevant document chunks at query time and grounds the LLM response in that context
- B
Increase the model temperature so the LLM explores more possible completions before answering policy questions
- C
Replace the LLM with a larger general-purpose model and remove document context, because larger models memorize more enterprise policy details
- D
Fine-tune the LLM only on historical support chats, without including the official policy and manual content in the inference workflow
Show answer and explanation
Correct answer: A
Explanation
The key issue is hallucination in a document-based question-answering workflow. Best practice for this type of LLM application is to ground generation using retrieval from trusted enterprise content, commonly referred to as retrieval-augmented generation (RAG). This allows the model to generate fluent answers while basing them on relevant passages from manuals, warranty terms, and policy documents. In practice, teams also often combine RAG with prompt instructions, citations, and lower-temperature settings for higher factual reliability. By contrast, simply increasing model size or temperature does not address grounding, and fine-tuning without retrieval does not reliably keep answers aligned with the latest official content. This aligns with common generative AI architecture guidance for enterprise LLM applications, including Snowflake-oriented patterns that pair vector search or semantic retrieval with LLM inference to improve answer accuracy.
- A. Correct.
Correct. Retrieval-augmented generation (RAG) is a standard pattern for reducing hallucinations in enterprise question-answering scenarios. Instead of relying only on the model's parametric knowledge, the application retrieves the most relevant passages from trusted source documents and provides them as context to the LLM at inference time. This grounds the response in authoritative data such as manuals and policy documents, which is especially important when answers must reflect current business rules.
- B. Incorrect.
Incorrect. Higher temperature increases randomness and creativity in token selection. That can be useful for brainstorming or marketing copy, but it generally makes factual consistency worse, not better. For support and policy answers, teams usually prefer lower temperature and stronger grounding in source content.
- C. Incorrect.
Incorrect. A larger model may improve general reasoning or language quality, but it does not guarantee accurate knowledge of a specific company's current warranty or return policies. Removing document context would typically increase the risk of unsupported answers because the model would have to rely on pretraining knowledge rather than the enterprise's source of truth.
- D. Incorrect.
Incorrect. Fine-tuning on historical support chats may teach tone and common response patterns, but if those chats contain mistakes, outdated policies, or incomplete information, the model can reproduce those issues. Fine-tuning alone also does not ensure that answers reflect the latest official documents at runtime. In this scenario, grounding responses in current source material is the more appropriate primary solution.