It's how an AI answers from your documents, not its training.
RAG means the system searches your actual, current documents for relevant information first, then uses that material to write an answer, instead of relying only on whatever the underlying model happened to learn during training.
It solves the out-of-date and made-it-up problems.
A model's training has a cutoff date and no knowledge of your internal documents at all. Retrieval fixes both: answers can reflect this week's policy update and point back to the actual source, instead of confidently guessing.
The search step matters more than the writing step.
How well the system finds the right passage, out of potentially thousands of documents, has more impact on answer quality than which AI model does the final writing. Most of the engineering work lives in this retrieval layer.
It's not the same as training a model on your data.
RAG doesn't change the model. It changes what the model is shown before it answers. That distinction matters for cost, speed, and how quickly the system can reflect a document that changed an hour ago.
Bring your workflow, your constraints, and the outcome you want to achieve.