Quantifying Retriever-Generator Alignment in RAG with Local Explanations
The RAG-E framework quantifies how well a retrieval system (RET) aligns with a generator (GEN) in Retrieval-Augmented Generation systems. It uses advanced attribution methods to find where components disagree, revealing significant misalignment. The paper introduces the Weighted Alignment of Retriever and Generator (WARG) metric as the first principled way to measure this alignment.