Grounding
Grounding is the practice of tying a language model’s answers to a specific, trusted source of information (documents, records or live system data). The model answers from what is provided rather than from what it can generate.
Also called: grounded generation, factual grounding
An ungrounded model answers from its training. A grounded one answers from the business: its price list, its service area, its calendar, its knowledge base. The information is supplied to the model at the moment it needs it, and the instructions tell it to rely on that material and say so when the material does not cover the question.
Grounding is the general principle; retrieval-augmented generation is the most common way of doing it for documents, and function calling is how it is done for live data such as an appointment slot or an order status.
It is the single most important property for an agent that speaks on a business’s behalf, because it turns "the AI said" into "the AI read out what we told it", which is the difference between a receptionist and a rumour.