You are discussing how to improve an LLM assistant that answers questions over internal documents. The team is considering better prompting, a RAG pipeline, and parameter-efficient fine-tuning because the current model is inconsistent on domain-specific questions.
How would you think about LLMs, RAG, and PEFT in this setting? When would you use each approach, how would you evaluate them, and how would you reduce hallucinations?
Choosing between prompt engineering, RAG, and PEFTDesigning LLM evaluation offline and onlineReducing hallucinations with grounding and refusalReasoning about enterprise retrieval quality