Retrieval vs fine-tuning
These mechanisms solve different problems and can be combined. One supplies selected information to a request; the other changes learned behavior through updates.
| Question | Retrieval | Fine-tuning |
|---|---|---|
| Mechanism | Find relevant material and supply it as context | Train parameters on a selected dataset |
| Best considered for | Fresh, inspectable, source-specific information | Repeated behavior, formatting, or task adaptation |
| Update route | Update the collection and retrieval index | Prepare examples and run further training |
| Primary risk | Missing, irrelevant, or malicious retrieved content | Poor data, overfitting, or unwanted behavior changes |
| Evidence | Can expose passages and source IDs | Does not automatically provide source provenance |
A support assistant can retrieve today’s policy while a fine-tuned model follows a consistent response format. Neither mechanism alone guarantees correctness.
Follow the evidence
Read the connected lessons for detailed mechanisms, assumptions, and primary references.
Retrieval: giving the model evidence
A retrieval system finds relevant material and places it where a language model can use it.
Supervised fine-tuning: demonstrations of behavior
Examples of good responses shift a pretrained model toward the behavior an application needs.
Evaluation: measuring the behavior you need
A useful score starts with a clear task, a representative test set, and a rule for deciding success.
Hallucinations: fluent text without support
A response can be well formed and plausible while inventing a fact, source, or event.