GraFine: Efficient Graph RAG with Retrieval-Time Refinement

A novel technique named GraFine has been introduced to enhance Graph RAG (Retrieval-Augmented Generation) in relation to corpus graphs. This method tackles two significant shortcomings found in current strategies: the lack of semantic awareness in graph expansion and the oversight of topology in pruning. GraFine operates through two distinct refinement phases: Semantic Proximity eXpansion (SPX), which focuses on adding nodes with semantic consideration, and a Graph Smoothing Reranker (GSR) that emphasizes graph-aware pruning. This innovative design integrates semantic node addition with efficient graph pruning, circumventing the slow retrieval and generation processes seen in earlier methods. Tests conducted on reference networks and text-rich knowledge graphs indicate that GraFine enhances both retrieval accuracy and generation quality while ensuring time efficiency. The paper can be accessed on arXiv with the identifier 2601.18579.