RAG, retrieval-augmented generation, is the architectural pattern where the agent searches a knowledge base (FAQs, policies, product docs) for documents relevant to the current question, then includes those documents in the model's context when generating the answer. The model is grounded in actual content rather than guessing.
Well-built RAG cuts hallucination dramatically on the slice of questions the knowledge base covers. Poorly-built RAG retrieves the wrong document, the model then ground-truths against the wrong source, which is worse than no RAG at all.
The competence axis is the retriever quality (does it find the right document for the question), the chunk strategy (how the knowledge is split into searchable units), and the answer-verification step (does the generated answer actually map to the retrieved content). Agentic RAG adds a second pass where the model evaluates whether the retrieved documents are sufficient and re-queries if not.
Vorel uses agentic RAG with vertical-specific retrievers, the clinic retriever is tuned differently from the auto-service one. Source documents stay in the customer's CMS so they own the knowledge.