Retrieval-Augmented Generation (RAG) is increasingly deployed in enterprise knowledge management, legal research, healthcare information support, and customer service. Yet production RAG systems are still routinely evaluated using single, narrow metrics that do not reflect the combined requirements of real-world deployment: retrieving the right evidence, citing sources accurately, generating answers that are faithful to that evidence, and operating within acceptable latency and cost budgets. This paper introduces CLEAR-RAG, a five-dimensional evaluation framework for production-grade RAG systems spanning Citation Quality (C), Latency and Cost (L), Evidence Faithfulness (E), Answer Relevance (A), and Retrieval Quality (R). The framework provides formal metric definitions, a 1–5 scoring rubric with domain-specific weighting presets, and an LLM-as-judge implementation that produces structured JSON scores. We describe a full reference implementation comprising both keyword and dense vector retrieval modes, a hybrid human-and-automated review workflow, and an error taxonomy covering nine failure categories. A controlled demonstration on 15 enterprise knowledge-base questions across two retrieval modes shows that CLEAR-RAG surfaces tradeoffs that are invisible to single-metric evaluation: keyword retrieval reaches the maximum Latency score (5.0) but trails on Retrieval Quality (2.8 average), while dense vector retrieval lifts Retrieval Quality to 3.7 at a modest latency cost (4.0). The framework is released as an open-source web application supporting both manual and automated evaluation workflows.
Md. Asadozzaman (Sun,) studied this question.