Grounding
Tying model claims to supplied sources — answer from the documents, not from vibes.
Grounding constrains generation to provided evidence: retrieved passages, database rows, tool results. A grounded system instructs the model to answer only from sources, cite them, and say “not found” otherwise — then verifies attribution. It is the working antidote to hallucination in production RAG systems, and the property most RAG evals actually measure.
Worked example: grounding forces the model to answer from provided source text (retrieved docs) rather than parametric memory, and citing the passage lets you verify it — the core defense against hallucination in RAG. Gotcha: it only helps if retrieval surfaced the right passage; with wrong or missing context a grounded model confidently answers from bad context — garbage in, confident garbage out.