R²Adapter
A lightweight, plug-and-play hybrid RAG routing and rewriting adapter that dynamically distributes queries between standard RAG and Graph RAG pipelines. It reduces graph retrieval overhead through intelligent routing and query rewriting, and is designed to be model-agnostic.
Timeline (last 90 days)
R²Adapter is a lightweight plug-and-play routing and rewriting adapter used to dynamically distribute queries between vanilla RAG and graph RAG.
On three multi-hop question answering benchmark datasets, R²Adapter can reduce graph RAG usage by up to 59% while maintaining answer accuracy comparable to full-pipeline graph RAG.
R²Adapter is model-agnostic and can be integrated into various vanilla RAG and graph RAG pipelines.
R²Adapter's routing function determines whether each query truly requires graph-structured reasoning, and only queries that can benefit from graph reasoning are routed to the graph RAG channel.
R²Adapter's rewriting function rewrites borderline queries to more clearly expose implicit multi-hop reasoning requirements, improving retrieval quality without the need for additional supervision signals.
Replacing heavyweight general-purpose large models with lightweight specialized components for routing decisions can achieve both speed and accuracy on intermediate decision tasks.
All Facts (6)
R²Adapter is a lightweight plug-and-play routing and rewriting adapter used to dynamically distribute queries between vanilla RAG and graph RAG.
50%UnverifiedOn three multi-hop question answering benchmark datasets, R²Adapter can reduce graph RAG usage by up to 59% while maintaining answer accuracy comparable to full-pipeline graph RAG.
50%UnverifiedR²Adapter is model-agnostic and can be integrated into various vanilla RAG and graph RAG pipelines.
50%UnverifiedR²Adapter's routing function determines whether each query truly requires graph-structured reasoning, and only queries that can benefit from graph reasoning are routed to the graph RAG channel.
50%UnverifiedR²Adapter's rewriting function rewrites borderline queries to more clearly expose implicit multi-hop reasoning requirements, improving retrieval quality without the need for additional supervision signals.
50%UnverifiedReplacing heavyweight general-purpose large models with lightweight specialized components for routing decisions can achieve both speed and accuracy on intermediate decision tasks.
50%