When do you choose RAG, fine-tuning or plain code?

When the model needs current knowledge from documents and sources, I start by retrieving those materials, i.e. RAG. When it's about a consistent style or behaviour across many examples, you can consider fine-tuning, i.e. additional training on chosen data. But first I check a simpler solution.

Fine-tuning isn't a convenient way to revoke access to a single file immediately. Plain code is better for exact calculations, comparisons and rules we already know. I don't need a model to check whether a number is positive. The methods can be combined, but each extra layer should solve a specific, measured problem.