Grounding
What is Grounding?
Tying a model's answers to verified sources such as your own documents, a database or live search, so it draws on real information instead of guessing. Grounded systems can cite where an answer came from, which makes them checkable. It is the main defence against hallucination, and retrieval-augmented generation is the common way to do it.
Why it matters
Grounding is often sold as the switch that stops a model making things up, but it only works as well as the sources behind it. Point a system at an outdated policy folder or let its retrieval pull the wrong document, and it will answer confidently from bad information, which can be harder to catch than an obvious hallucination because it looks sourced. The real value is accountability. When every answer carries a link back to where it came from, a person can check it, an auditor can trace it, and a wrong answer can be traced to a fixable source rather than a mysterious model quirk. That traceability is what makes AI usable where being wrong has consequences.
In practice
A lender’s internal assistant answers staff questions about credit rules and cites the exact clause behind each reply. When a regulator later asks how a decision was reached, the trail is already there. The same setup surfaces a subtler problem: one answer cites a policy document that was superseded months ago, showing that the source library, not the model, needed maintaining. Grounding turned a hidden data problem into a visible one.
