Grounding
- Context
- Standard term
Grounding means tying an AI's answer to evidence outside the model, such as a document, a database record, a calculation, or a tool's output, so each claim can be traced to what supports it.
Left alone, a model answers from its training data, a blend of many texts with no record of which said what. A grounded answer is written with evidence in view and cites the part it used. Finding evidence and writing can fail separately, so an answer can cite a real passage that says something else.
Where you’ll meet it
Your company's HR assistant in the work chat tells you that five unused vacation days carry over into next year. It sounds sure and shows no source. You're about to book a trip around those days.
What to do
Ask it to quote the handbook passage behind the answer, then find that passage in the handbook and read it yourself. If it says something different, or the assistant can't point to one, check with HR before you book.
Easy to mix up
RAG is one method: the system fetches documents and answers from them. Grounding can also come from a file you pasted or a calculation's result, with no search involved.
Check yourself
Your assistant answers a question about your health plan from last year's benefits guide and quotes the exact page. The plan changed in January. Is the answer grounded?
Go deeperAutomated support checks have known gaps, and citation is usually selective
A common check splits an answer into claims and asks whether the cited passage entails each, meaning the claim follows from it. A second model, or a smaller one trained only to say yes or no, does the asking, and benchmarks call the result attribution. Such checks catch a mismatched citation and can miss a claim that combines two passages.
Citing every sentence slows answers and can crowd out general knowledge that needed no source. A common choice is to require evidence for the amounts, dates, and rules a decision rests on, and for facts whose sources change often. Creative tasks, such as naming a product, have nothing to cite and are better judged against examples and constraints.