Grounding
Making a model answer from a specific set of documents you give it rather than from memory.
A language model answering from training alone is working from a blurry recollection of a lot of text, which is where confident nonsense comes from. Grounding narrows the job: you hand the model specific sources and ask it to answer from those, usually with citations pointing back to the passage it used. Notebook style tools work this way, and so does most enterprise AI search.
It helps a great deal and it is not a guarantee. A grounded model can still summarise a passage wrongly, stitch two unrelated bits together, or cite something that does not quite say what it claims. The useful habit is to treat the citation as the product: if an answer matters, click through and read the sentence it came from. Grounding also changes what a tool is for, since a grounded assistant is only as good as the sources you gave it.