Context Utilization
The extent to which a RAG system’s response actually incorporates information from the retrieved context.
High relevance + low utilization = the retriever worked but the generator ignored it. This diagnostic combination points to a generator problem, not a retrieval problem.
Utilization answers: “The right information was available, did the model use it?” This is distinct from adherence, which asks whether the model strayed beyond the context. A response can be perfectly adherent (no hallucination) while having low utilization (ignoring relevant retrieved content).
The metric emerged because existing frameworks focused on faithfulness to context but not on completeness of use of that context. You can be faithful to what you cite while citing too little.