SnowPro Specialty: Gen AI Question 55
Single answerRAG use casesA financial services company uses a retrieval-augmented generation (RAG) assistant built on Snowflake to answer employee questions about internal compliance policies. The assistant currently retrieves policy documents by keyword match, and users report that answers sometimes cite outdated policy versions or miss relevant sections when different terminology is used. The team wants to improve answer grounding and relevance without retraining the LLM. Which approach is the BEST fit for this RAG use case?
- A
Replace retrieval with a larger language model so the model can infer the latest policy content from its pretraining
- B
Implement semantic retrieval over chunked policy documents with metadata filters for document version and effective date before passing context to the model
- C
Fine-tune the language model on all policy documents and remove citations because the model will internalize the compliance rules
- D
Increase the temperature of the generation model so it explores a wider range of possible policy interpretations
Show answer and explanation
Correct answer: B
Explanation
The best answer is to improve the retrieval stage of the RAG pipeline rather than relying on model size, fine-tuning, or generation randomness. In enterprise RAG use cases, especially for compliance and policy question answering, best practices include chunking source documents, using semantic search to improve matching beyond keyword overlap, and applying metadata filters to restrict results to current, approved, or in-force documents. This keeps answers grounded in authoritative sources and reduces the risk of stale or irrelevant context. Snowflake guidance for Gen AI and Cortex-related retrieval workflows emphasizes using high-quality retrieved context and structured filtering to improve grounded generation, which is the core value of RAG when documents change over time.
- A. Incorrect.
This is incorrect because relying on a larger general-purpose LLM does not guarantee access to the company's latest internal compliance policies. In a RAG pattern, retrieval supplies current, authoritative context at inference time. Pretrained model knowledge can be stale and may hallucinate details, which is especially risky for compliance content.
- B. Correct.
This is correct because semantic retrieval helps match meaning rather than exact keywords, improving recall when users and documents use different terminology. Chunking documents improves retrieval granularity so the model gets the most relevant sections instead of entire long documents. Applying metadata filters such as version, effective date, or policy status helps ensure that only current and valid policy content is retrieved, which directly addresses the issue of outdated citations. This is a standard best practice for enterprise RAG systems.
- C. Incorrect.
This is incorrect because fine-tuning is not the best solution for keeping answers grounded in frequently updated source documents. Fine-tuning can bake information into model weights, making updates harder and reducing transparency. Removing citations is also a poor choice for compliance use cases, where traceability to source documents is important for trust and auditability.
- D. Incorrect.
This is incorrect because temperature affects randomness in generation, not retrieval quality or source freshness. Raising temperature may actually make answers less consistent and less suitable for regulated policy guidance. The reported problem is about finding the right supporting documents, so the retrieval pipeline should be improved rather than making generation more creative.