To address the challenge of context attribution, which involves attributing responses from Retrieval-Augmented Generation (RAG) models to specific context segments, this study proposes Attribute Response to Context (ARC-JSD), a novel methodology leveraging Jensen-Shannon Divergence. ARC-JSD efficiently and accurately identifies essential context sentences without additional fine-tuning, gradient calculations, or surrogate modeling. It demonstrates superior accuracy and computational efficiency over existing surrogate-based methods on various RAG benchmarks, and analyzes specific attention heads and MLP layers to provide insight into the model's internal workings.
