What is grounding?
Grounding is the practice of anchoring a generative AI model’s output to verified sources that are supplied at the time of the request. Instead of answering from whatever it absorbed during training, the model receives the relevant policy text, the current record from the database, or the result of a tool call, and is instructed to answer only from that material. The sources are usually shown to the user alongside the answer so they can check it.
Retrieval-augmented generation (RAG) is the most common grounding technique for documents. Grounding also covers structured data through text-to-SQL and function calling, where the model asks a system of record for a figure rather than recalling one. A grounded assistant that cannot find support for an answer is expected to say so.
Why grounding matters
A model’s training data has a knowledge cutoff and contains nothing about a specific company’s contracts, prices, or procedures. Without grounding, the model fills those gaps with plausible guesses, which is the mechanism behind hallucination. Grounding converts the model from a source of facts into a reader and writer of facts that live elsewhere. That shift makes answers auditable, because a reviewer can trace each statement to the passage or record it came from.
Grounding also keeps answers current. When the policy changes, the document store is updated and the next answer reflects it. No retraining is needed.
Grounding in practice
A county planning department runs an assistant for residents asking about zoning and permit rules. Each question is answered from the current ordinance text and fee schedule, with the section number cited. When a resident asks about a situation the ordinance does not address, the assistant says that it found no applicable rule and offers the planning desk phone number. Staff review a sample of conversations each week and have found that the cited sections make it quick to spot a wrong answer.
What to watch for
- Grounding is only as good as the retrieval step, so a wrong or outdated document produces a confidently wrong answer with a citation attached.
- Instruct the model explicitly to decline when the sources do not contain the answer, and test that behavior with questions you know are unanswerable.
- Pull numbers, dates, and identifiers directly from the system of record rather than letting the model restate them from a document.
- Check that citations point to text that supports the claim, because a model can attach a real reference to a sentence it invented.