What is groundedness and how do we measure it?

Groundedness is the extent to which an LLM output is supported by provided source material. Measure with an LLM-as-judge grader that scores each claim in the answer against the retrieved context: fully supported, partially supported, unsupported. Track the unsupported-claim rate as a production metric — anything above 5% in RAG systems is a red flag.