Written and reviewed by Manish Kulkarni · About the author
If you want an AI model to behave differently for your use case, you have three main options, ordered here from simplest to most involved: change what you ask (prompting), give it your own documents to look at (retrieval), or change the model itself (fine-tuning). They solve different problems.
Prompting means giving instructions, context or examples with each request. The model's weights never change. This covers basic prompts, detailed system instructions and example-based prompting (see Prompting Fundamentals).
Often the real problem is that the model doesn't know your information, such as your product manual or policies. Retrieval (often called RAG, retrieval-augmented generation) solves this by finding the relevant passages and including them in the prompt for each question. The model's weights still don't change, and your information can be updated any time by changing the documents.
Fine-tuning continues training an existing model on your own dataset, actually adjusting its weights (see Training vs Inference). It is best for making a model consistently follow a particular style, format or narrow task, without repeating long instructions every time.
Fine-tuning is better at changing how a model responds than at teaching it new facts reliably. If your goal is "answer questions using our latest documentation", retrieval is usually the better fit, because the documents can change without retraining anything.
Generally more expensive and complex than prompting, requiring quality training data and technical setup — most projects should try thorough prompting first before considering fine-tuning.
No — it typically makes a model better at a specific narrow task or style, sometimes at the cost of general flexibility elsewhere.
Not all providers offer fine-tuning access, and availability varies by model and provider — check the specific provider's documentation for what's supported.