Groundedness Testing
Groundedness testing checks whether an AI-generated answer is actually supported by the source material it was given — retrieved documents, provided context — rather than drifting into the model's own unverified internal knowledge. It's distinct from hallucination testing, which checks factual correctness regardless of source.
A RAG system can retrieve exactly the right documents and still produce an ungrounded answer, because nothing forces the model to stick to what it was handed — it can blend in outside knowledge, over-generalize from one line of a source, or state something the source only implies. Groundedness testing isolates this specific failure: given the retrieved context, is every claim in the answer traceable back to it, sentence by sentence?
The standard technique is a groundedness scorer (often another LLM acting as judge) that takes the answer and the source context and flags any claim not supported by it — separately from checking whether the claim is true. This distinction matters: an answer can be perfectly grounded in a source document that's itself wrong, and a hallucination can be accidentally true. Teams building RAG features track groundedness as its own metric precisely because it's the one that's actually under the system's control — you can't guarantee source documents are correct, but you can guarantee the model only uses them.
