What is a Grounding (AI)?

A language model's parametric knowledge is frozen at training time and can be wrong, outdated, or simply absent for a given fact. Grounding closes that gap by feeding the model retrieved passages, records, or API responses as part of its context window, so its answer is computed from evidence that exists right now rather than from a statistical guess. Techniques range from retrieval-augmented generation over a vector store to direct tool calls that pull a live number from a CRM or spreadsheet. Grounding matters most wherever a wrong answer has consequences: compliance questions, financial figures, customer-specific data, or anything time-sensitive. It also underpins citation — a grounded answer can point back to the exact source passage, which lets a human verify the claim instead of trusting the model on faith. Ungrounded generation and grounded generation can look identical in tone; the difference only shows up when you check the facts.