Better prompt, retrieval or fine-tuning: try them in that order

When output is not good enough there are three levers, and the cost and reversibility differ by an order of magnitude at each step. Start with the prompt and the context. Most disappointing output comes from an instruction that is ambiguous, a missing example of what good looks like, or a task that is too large for one step. Splitting a job into two calls with a check between them fixes an enormous share of problems. Changes here are instant and free to undo. Then retrieval. If the failure is the model not knowing something — your documents, your catalogue, your policies, anything recent — the answer is to put the right information in the context, not to train it in. Retrieval keeps the knowledge editable: correcting a fact means fixing a document, not retraining. Most business problems described as "the model does not know our business" are retrieval problems. Fine-tuning earns its place for a narrower thing: consistent format, a specific tone, or a specialised classification where the examples are plentiful and the task is stable. It does not reliably teach facts, it needs a real dataset, it has to be redone when the base model changes, and it makes the system harder to move. A practical discipline: before doing anything expensive, write down what specifically is wrong with the current output on ten real cases. Very often the description of the problem contains its own answer — and it is almost never "the model needs training". This article is general information, not legal, tax or financial advice. Rules change and every deal is different — check your own case with a licensed professional. Alberto Zaltzberg — Adonait · adonait.com