RAG: giving an assistant the right source material

Two connected stages

A RAG system first finds relevant material, then supplies it to a model that produces an answer. This can connect an assistant to a changing document collection without retraining the model for every update. It does not ensure the retrieved passages are current, complete or correctly interpreted.

Trace one answer backwards

Suppose an employee asks about parental leave. A useful system must find the policy for the right location and employment situation, then preserve its conditions in the response. If the wrong document is retrieved, a beautifully written summary may still be wrong. If the correct document is retrieved, the model can still omit an exception.

Improve the weakest step

Test retrieval and answer quality separately. Record the query, passages, document versions and final response. Add questions that have no supported answer and check whether the assistant admits the gap. Keep permissions attached to the underlying sources. A practical first RAG project is a small, well-maintained collection with a clear owner; an enormous uncurated document dump makes errors harder to diagnose.

Sources & notes

AI-assisted editorial content checked against the linked sources.

Lewis and colleagues: Retrieval-Augmented Generation

Sources reviewed for the September 2026 launch edition.