R$^{2}$Adapter: A Routing and Rewriting Adapter for Efficient Hybrid RAG

A question-answering system need not use its most elaborate search method for every question. R²Adapter chooses between retrieving text directly and searching a graph of relationships, which can help when an answer requires connecting several facts. It also rewrites uncertain graph-bound questions to make those connections easier to retrieve. Across three question-answering benchmarks requiring linked reasoning, the authors report up to 59% less use of graph-based retrieval while maintaining comparable answer accuracy. This supports routing questions according to their needs; the reported reduction measures use of the graph pathway, rather than an equivalent percentage reduction in total cost or response time. Study abstract.