To make a foundation model fit a business, start with the cheapest option that works: prompt engineering first, then retrieval-augmented generation (RAG) for missing knowledge, then fine-tuning for missing behavior. The AWS AI Practitioner (AIF-C01) exam tests this as a cost ladder that also includes model distillation and pre-training (objective 3.1.5).
Last updated October 2026.
Is the model missing knowledge or missing behavior?
Missing knowledge means facts: this quarter's price list, internal policies, product manuals. Put those in the prompt or retrieve them with RAG. Missing behavior means a skill, style or format the model cannot produce reliably however you ask: a house writing style, a specialized classification, a domain vocabulary used consistently. That calls for training.
Most wrong answers on this topic come from mixing the two up, usually by choosing fine-tuning to teach facts that change every week.
Instructions and a few examples go in the prompt; the model's weights never change. There is no training cost, only extra prompt tokens on every call. It is best for format, tone and simple rules a handful of examples can show, and changes take minutes. The limit is the context window, and long prompts cost tokens on every request.
Relevant passages are retrieved from your documents and added to the prompt. On AWS this is Amazon Bedrock Knowledge Bases with a vector store. You pay for embeddings, the vector store and the retrieved tokens, but there is no model training. Updating knowledge is a data sync, not a retrain, and answers can cite their sources. It adds facts, not skills.
Further training of an existing model on labeled prompt and response pairs, so its weights change. You pay for data preparation and a training job up front, then for hosting the custom model. It is the right tool for consistent new behavior, but it needs hundreds to thousands of good examples, and new facts mean another training run.
A large teacher model answers your prompts, and those answers train a smaller student model. Amazon Bedrock Model Distillation automates this. AWS states distilled models are up to 500% faster and up to 75% less expensive than the original models, with less than 2% accuracy loss for use cases like RAG. Use it when a big model is accurate but too slow or costly at volume.
Training a foundation model from scratch on a huge body of mostly unlabeled data. It needs massive data, large accelerator clusters for weeks or months, and a specialist team. On the exam it is almost never the answer for a single business; continued pre-training on unlabeled domain text is the cheaper middle ground.
Objective 3.1.5 from the official AIF-C01 exam guide, version 1.1, which added distillation. Distillation figures from the Amazon Bedrock Model Distillation product page, checked 23 October 2026.
| Option | Up-front cost | What it changes |
|---|---|---|
| Prompt engineering | None | Format and instructions |
| RAG | Low | Knowledge |
| Fine-tuning | Medium | Behavior and skill |
| Distillation | Medium to high | Big-model skill in a small model |
| Pre-training | Extreme | Everything |
Per-request cost runs the other way at the top: prompting and RAG add tokens to every call, while a fine-tuned or distilled model adds hosting, and distillation's small model is the cheapest to run at volume.
The common trap: a scenario says the model gives outdated answers about frequently changing products and offers fine-tuning as an option. Fine-tuning would be out of date again after the next change. RAG keeps answers current with a sync.
The rungs are not exclusive. A fine-tuned model that has learned a house style can still use RAG for current facts, and every option still needs a well-written prompt. On the exam, though, the question usually asks for the first or most cost-effective step, and that is the lowest rung that closes the gap.
RAG adds knowledge at request time by retrieving relevant documents into the prompt, so the model itself does not change. Fine-tuning changes the model's weights by training it on labeled examples, which teaches new behavior such as a style or specialized task.
Use RAG when the model needs private or frequently changing information, especially if answers must cite sources. On AWS that means Amazon Bedrock Knowledge Bases; updating knowledge is a data sync rather than a new training job.
Distillation trains a smaller student model on answers generated by a larger teacher model for your prompts. AWS states distilled models are up to 500% faster and up to 75% less expensive than the originals, with less than 2% accuracy loss for use cases like RAG.
Prompt engineering, also called in-context learning. It needs no training and no extra hosting, and the only cost is the additional prompt tokens on each request.
Unit 3 of the Certified AIF-C01 course covers model choice, RAG, vector stores, the customization ladder, prompting and evaluation, with exam-style questions in every chapter.
Open the AIF-C01 course →