Fine-Tuning versus Prompting versus RAG

Three techniques dominate practical AI engineering, and they are frequently confused. Each solves a different problem.

Prompting

Change the instructions. This is the cheapest and fastest lever, and it addresses behaviour: tone, format, refusal patterns, output structure. If the model has the knowledge but expresses it wrongly, prompting is the answer.

Retrieval

Supply the missing knowledge at request time from a searchable index. This addresses facts: internal policies, product catalogues, recent documents. It updates instantly when the source does, and it produces citations.

Fine-tuning

Continue training on your own examples to adjust the model's weights. This addresses style and format consistency at scale, or behaviour that must hold under many different inputs. It does not reliably teach new facts, and it costs money every time the base model is updated.

A practical order of operations

1. Write a clear prompt with examples of the output you want.
2. Add retrieval if the model lacks the knowledge.
3. Fine-tune only when consistent behaviour cannot be achieved any other way.

Teams that follow this order spend far less, and they keep the ability to switch base models when a better one appears.