SnowPro Specialty: Gen AI Question 56
Single answerRAG use casesA company is building a customer-support assistant in Snowflake to answer questions from product manuals, policy PDFs, and release notes. Early testing shows the model gives fluent but occasionally outdated answers because the base model was trained before the latest product changes. The team wants to improve factual accuracy without fine-tuning the model every time documents change. Which approach is the MOST appropriate?
- A
Implement a RAG pipeline that retrieves relevant document chunks from the latest enterprise content at query time and provides them as grounded context to the LLM before generation.
- B
Increase the temperature of the LLM so it produces more diverse answers and is less likely to repeat outdated training data.
- C
Convert all source documents into a single larger prompt and send the entire corpus with every user question so the model has maximum context.
- D
Rely on prompt instructions alone, telling the model to avoid outdated information and answer only from its pretraining knowledge.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use RAG: retrieve relevant content from the organization's latest documents and include that retrieved context in the model prompt at inference time. This pattern is especially useful when source content changes frequently and the goal is to improve factual grounding without repeatedly fine-tuning a model. In practice, effective RAG systems chunk documents, create embeddings, store or index them for similarity search, retrieve top-matching chunks for a user query, and use those chunks to ground the response. This aligns with common Snowflake GenAI best practices around using retrieval for enterprise knowledge grounding rather than depending solely on model pretraining. Key reasons the other options are weaker: sampling parameters such as temperature do not address stale knowledge, passing the full corpus is inefficient and often infeasible, and prompt-only approaches do not provide access to updated facts.
- A. Correct.
Correct. Retrieval-augmented generation (RAG) is designed for this exact scenario: the model's parametric knowledge may be stale, but the organization has newer source documents. By retrieving semantically relevant chunks at inference time and passing them into the prompt, the application grounds responses in current enterprise data without requiring frequent fine-tuning. This is the practical pattern for document Q&A over changing corpora such as manuals, policies, and release notes.
- B. Incorrect.
Incorrect. Temperature controls randomness and creativity in generation, not factual freshness. Raising temperature usually makes outputs less deterministic and can increase variation, but it does not solve the core issue that the model lacks access to updated documents at answer time.
- C. Incorrect.
Incorrect. Sending the entire corpus for every question is not a scalable or effective RAG design. It increases cost and latency, can exceed context-window limits, and may reduce answer quality because the model must sift through too much irrelevant information. Best practice is to retrieve only the most relevant chunks for the specific query.
- D. Incorrect.
Incorrect. Prompt instructions can influence behavior, but they cannot give the model access to new facts that are not in its context window. Telling the model to avoid outdated information does not ground it in the latest manuals or release notes. This option reflects the misconception that prompting alone can replace retrieval over current enterprise data.