RAG Is Useful, But It Is Not Magic

Retrieval-augmented generation sounds like the solution to AI hallucination.

Connect the model to your documents. Ask questions. Get grounded answers.

Useful, yes. Magic, no.

RAG systems still fail in very ordinary ways. They retrieve the wrong document. They retrieve the right document but the wrong section. They miss context from another system. They over-trust outdated content. They summarise conflicting documents as if there is no conflict.

The hardest part of RAG is not the model. It is the information estate.

Are documents current?

Are permissions correct?

Are naming conventions sane?

Are versions controlled?

Are business rules written down?

Are exceptions captured?

Most companies do not have an AI problem. They have a knowledge management problem that AI exposes.

RAG can be excellent when it is built on clean, governed, structured information. When it is built on years of document clutter, it becomes a confident search assistant with a memory problem.

Still worth building. Just not worth worshipping.