AI · ArchitectureSeptember 14, 20264 min · updated September 21, 2026
RAG, fine-tuning or a custom model: an eight-step framework
Rephrased by Daillac
Source: AWS ↗D
In brief
- AWS organizes AI customization into eight levels of increasing complexity.
- The proposed principle is to start with the simplest approach that meets the criteria.
- RAG and fine-tuning solve different problems and can be combined.
Start with the need, not the technique
The framework moves from direct model use through structured prompting, RAG, distillation, fine-tuning, continued pre-training and fully custom models. AWS notes that most workloads do not need to move beyond the early levels.
8 steps
make up the proposed spectrum, from prompting to a fully custom model.
Source : AWS
What this changes for an organization
RAG fits knowledge that changes and needs citations. Fine-tuning is better suited to stable behaviour, formatting or specialization. Before either, a representative evaluation set should define expected quality.
Four controls to put in place
- Write success criteria and an evaluation set.
- Measure a base model with structured prompting.
- Add RAG when the gap is knowledge.
- Consider fine-tuning when the gap is stable, repeated behaviour.
What this announcement does not establish
The eight-step framework is a decision guide rather than an automatic business case. Costs vary with volume, context size, update frequency, data quality and evaluation effort. A simple approach can still become expensive when every request carries too many documents.
TableDecision framework · shareable block
| Avoid | Do | |
|---|---|---|
| 01 | AWS organizes AI customization into eight levels of increasing complexity. | Write success criteria and an evaluation set. |
| 02 | The proposed principle is to start with the simplest approach that meets the criteria. | Measure a base model with structured prompting. |
| 03 | RAG and fine-tuning solve different problems and can be combined. | Add RAG when the gap is knowledge. |
Practical questions
What exactly does the primary source announce?+
The framework moves from direct model use through structured prompting, RAG, distillation, fine-tuning, continued pre-training and fully custom models. AWS notes that most workloads do not need to move beyond the early levels.
What is a reasonable first action?+
Write success criteria and an evaluation set. Measure a base model with structured prompting.
Which limitation should remain in view?+
The eight-step framework is a decision guide rather than an automatic business case. Costs vary with volume, context size, update frequency, data quality and evaluation effort. A simple approach can still become expensive when every request carries too many documents.
How should implementation be monitored?+
Add RAG when the gap is knowledge. Consider fine-tuning when the gap is stable, repeated behaviour.
Turn this news into a concrete decision
DAILLAC can define the architecture, controls and measurements that fit your organization.
Sources & method
Article written from two primary sources, verified on September 21, 2026, then contextualized for Québec and Canadian organizations.
Read the original source: AWS ↗