Tying a model's answers to specific, trusted information sources instead of relying only on its general knowledge.
Grounding means a model's answer is built on real information you provide: your company documents, database records or a tool's result. A grounded model answers according to the attached return policy instead of guessing what the policy might be. That makes the answer checkable.
It is one of the most important ways to reduce hallucinations and is essential in any app that gives customers information they rely on: prices, appointments, terms. The most common methods are RAG and tool calling to fetch live data. Practical tip: instruct the assistant to answer only from the provided sources and cite them, and to say clearly when it cannot find the answer. Then review a sample of replies regularly to confirm it actually follows that rule.