Retrieval-augmented generation (RAG)

When a question is asked, the system first searches a knowledge base – such as manuals, contracts or the intranet – for the most relevant passages. These are passed to the model together with the question, and the model formulates an answer with citations.

RAG is the most common way to connect AI with company knowledge: the knowledge stays current and controllable, hallucinations become rarer, and no fine-tuning is required. Quality depends heavily on good data preparation and search.