Evaluating RAG with LLM as a Judge

Mistral proposes judge LLMs scoring generator responses on numerical, binary or qualitative scales, then taking weighted averages across evaluation datasets to assess RAG systems.