Grounding is the highest-leverage single intervention against hallucination, because it replaces recall with reading.
It reduces the problem rather than eliminating it. A grounded model can still overstate what a source says, and when retrieval returns something irrelevant but plausible, grounding makes the answer worse than abstaining would have been.