Different changes
RAG retrieves source material for a request. Fine-tuning changes model behavior through training examples. A new policy that must be cited is a different requirement from a stable classification style. Neither method removes the need to evaluate outputs or enforce who may access the data.
For a fictional repair supplier, the warranty document changes while the response format stays fixed. Start by retrieving the approved revision and checking citation membership. Training examples alone do not establish that a response used the current policy. If repeated routing decisions remain wrong despite clear instructions and representative examples, evaluate that separate behavioral problem before considering training.
Evidence before a choice
Keep a frozen set of development and held-out cases. Check source freshness, unsupported answers and access boundaries separately from formatting. A method that improves one dimension may leave another unchanged. You can combine retrieval and training, but that does not make their costs or failure modes interchangeable.
The OpenAI optimization guide, checked November 14, 2026, labels fine-tuning workflows legacy and says its fine-tuning platform is unavailable to new users. That is a platform availability fact, not a statement that the method has ceased to exist. Check the current provider route before planning any training job.
Practice scope
Build the local RAG project and its evaluation split. The notebook runs original documents and prepared replies. It does not train weights, upload a dataset or call a live model. Use the exercise to test application logic, then evaluate any live architecture in an approved environment.