Grounding defect rate (GDR) measures the share of generated responses that are ungrounded: ungrounded responses divided by total generated outputs. It is a response-level metric, not a count of unsupported claims within each response. Track it when an assistant may invent product specifications or shipping estimates instead of using available information. section 9.1.1 section 9.0
To measure GDR:
- Validate generic test responses against the knowledge base. Semantic similarity, exact matching of specific values, and rule-based extraction of fields such as prices or dates can flag discrepancies and unsupported claims. section 9.1.1
- Have domain experts review adversarial edge cases, where automated checks may miss nuance. Classify flagged responses as accurate, partially accurate, or hallucinated. section 9.1.1
- Report GDR separately for generic and adversarial tests, then track it over time to catch regressions after prompt or model updates. Break results down by topic to locate weak areas. section 9.1.1
Read the breakdown, not just the average
A system might have 8% GDR on generic tests but 25% on adversarial tests, pointing to greater difficulty with edge cases. A topic breakdown might show 3% for product information and 15% for shipping policies, giving the team a clearer improvement target. section 9.1.1
GDR does not express how harmful an error is. Use a separate hallucination severity score to distinguish minor discrepancies from critical inaccuracies. Grounding practices—retrieving relevant documents with RAG, providing citations, and allowing an “I don’t know” response—address the failures GDR is meant to reveal; continued measurement checks whether those practices hold up. section 9.1.1 section 11.6.6
