When businesses want a large language model to "know" their information, there are two main options: retrieval-augmented generation (RAG) and fine-tuning.
Retrieval-augmented generation
RAG searches your documents at question time and gives the most relevant passages to the model. The model's answers can cite sources, stay current when documents change, and respect user permissions. For most knowledge assistants, RAG is the right starting point.
Fine-tuning
Fine-tuning further trains a model on examples, so it learns a style, a format or a narrow task. It is useful for consistent outputs such as classification. It is not a good way to store facts that change.
Our rule of thumb
| Need | Best fit |
|---|---|
| Answer questions from policies and documents | RAG |
| Consistent tone or output format | Prompting first, then fine-tuning |
| Classify high volumes of text cheaply | Fine-tuned small model |
| Up-to-date facts with citations | RAG |
Whichever approach you choose, invest in an evaluation set early. It is the only reliable way to know whether a change made things better.