How Fine-Grained Should a RAG Benchmark Be? A Hierarchical Framework for Synthetic Question Generation

Chase M. Fensore, Kaustubh Dhole, Jason Fan +2

Evaluating retrieval-augmented generation (RAG) systems requires benchmarks that capture diverse question characteristics, yet practitioners lack empirical guidance on which dimens…