SnowPro Specialty: Gen AI Question 284
Single answerImplementation methodsA retail company is building an internal product-support assistant in Snowflake. The assistant must answer questions using the company's warranty guides and troubleshooting manuals, while ensuring responses are grounded in the latest internal documents rather than only a model's pretrained knowledge. The team wants an implementation method that minimizes model retraining effort and can be updated quickly as documents change. Which approach is the MOST appropriate?
- A
Fine-tune a large language model every time warranty guides are updated so the model memorizes the latest documents
- B
Implement retrieval-augmented generation (RAG) by storing document embeddings, retrieving relevant document chunks at query time, and passing them as context to the model
- C
Use prompt engineering alone with a detailed system prompt that describes the warranty policies, without connecting to any document store
- D
Train a custom classification model to assign each user question to a product category, then return a static template response for that category
Show answer and explanation
Correct answer: B
Explanation
The best implementation method is retrieval-augmented generation (RAG). In enterprise Gen AI solutions, RAG is commonly preferred when answers must be based on proprietary or frequently changing data, because it separates knowledge updates from model training. Instead of retraining the model, teams update the source documents and retrieval index. This improves maintainability, reduces cost, and helps produce responses grounded in authoritative content. In Snowflake-based implementations, this pattern aligns with best practices around using vector embeddings and similarity search to retrieve relevant context before LLM inference. Fine-tuning is more appropriate when the goal is to adapt model behavior, tone, or task performance, not to keep pace with frequently changing documents. Prompt engineering is useful but complementary, not a substitute for retrieval when current enterprise knowledge must be included.
- A. Incorrect.
This is incorrect because repeatedly fine-tuning a model for document updates is typically expensive, operationally heavy, and slower to maintain. For rapidly changing enterprise knowledge bases, fine-tuning is usually not the first choice when the goal is to keep answers grounded in current documents. Fine-tuning can help with style, task specialization, or behavior, but it is not the most efficient implementation method for continuously changing source content.
- B. Correct.
This is correct because retrieval-augmented generation (RAG) is designed for exactly this scenario: grounding model responses in external, current enterprise content. By chunking documents, generating embeddings, retrieving the most relevant passages at inference time, and providing them to the model in the prompt, the assistant can answer using up-to-date internal documentation without retraining the base model whenever content changes.
- C. Incorrect.
This is incorrect because prompt engineering alone does not provide reliable access to the latest internal documents. A strong system prompt can improve instructions and response format, but it cannot dynamically inject current source material unless that material is also provided at runtime. This option reflects the common misconception that a better prompt can replace a retrieval layer for knowledge-grounded applications.
- D. Incorrect.
This is incorrect because classification plus static templates may work for simple routing or FAQ-style automation, but it does not meet the requirement to generate grounded answers from detailed manuals and warranty guides. It also lacks the flexibility needed for nuanced, document-based responses and follow-up questions.