An open-book answer
If you ask an AI assistant about your company’s leave policy, the app can search policy documents you are allowed to see. It sends relevant passages with your question to a language model, which uses them to draft an answer. This is RAG.
The documents stay separate from the model, so they can be updated without retraining it. The answer can also link back to its sources.
When the search falls short
The answer can still be wrong if the app finds outdated or irrelevant material, misses an important document, or the model misreads a passage. The app must respect document permissions, and citations should support the claims they accompany. RAG can reduce made-up claims, but cannot prevent them all.