RAG vs fine-tuning

Fine-tuning adjusts the weights of an existing LLM on a task-specific dataset to bake in behavior - expensive (training infra, MLE weeks), better at adapting style, format and behavior, but doesn't add new knowledge reliably. RAG attaches an external retriever (BM25, dense embeddings) and injects passages into the prompt - cheap to update (just swap the corpus), gives attribution, but is bounded by context window, retrieval quality, and budget. Default to RAG when the data is large, evolving, or needs attribution (catalog QA, internal docs, support); fine-tune when the task requires consistent voice/format and the dataset is small and frozen. Combine both: fine-tune the base for tone and safety; RAG for facts. Decision tree: (1) is knowledge evolving? RAG. (2) is voice/style the bottleneck? FT. (3) compute budget? RAG cheaper; FT when you can amortize across millions of queries.