A List-Aware Re-Ranking Model with Multi-Granularity Relevance Feature Fusion

In the retrieval stage of Retrieval-Augmented Generation (RAG), a subset of documents or text chunks is typically first retrieved from the database. These retrieved chunks are then reranked, and the top-ranked ones are selected as the final output. Most existing ranking methods focus solely on the similarity between the query and individual relevant documents while neglecting the interrelationships among the relevant documents. In this study, we propose a list-level reranking model based on multi-granularity relevance features fusion (MGRF), which explicitly incorporates the relationships among retrieved text chunks within the candidate list to improve ranking. Our model incorporates multi-granularity relevance features between the query and texts, as well as among the text chunks themselves, to more effectively model their interrelationships. Our model is extensively evaluated on public datasets and achieves superior results, demonstrating the effectiveness of our approach.