Optimizing Information Retrieval in RAG through Intelligent Reranking and Follow-Up Query Predictions

Abstract Reranking is a crucial process in Retrieval-Augmented Generation (RAG) systems as it significantly impacts the quality and the relevance of retrieved knowledge chunks. Conventional reranking models usually prioritize semantic similarity and matching accuracy between user queries and knowledge base embeddings. This causes them to often lack the ability to dynamically adapt to the evolving context of user interactions. In this paper, we propose a novel reranking framework designed to fill this gap and enhance retrieval in RAG systems by incorporating LLM-generated predicted follow-up queries coupled with the initial user query to better capture the evolving user intent. Our model leverages a fine-tuned weighting mechanism to balance the embeddings of the initial query and predicted follow-up queries, enabling context-aware reranking of knowledge chunks. The proposed approach tackles critical challenges, including enhancing personalization, scalability and ensuring relevance in scenarios where user queries are dynamic and context dependent. It addresses the need for adaptive retrieval mechanisms that can effectively handle evolving user intent and context to improve the quality of retrieved information. Evaluation on two benchmark datasets demonstrates that our reranking framework improves retrieval quality, effectively integrating user intent prediction to optimize the RAG process. Our results highlight the potential of embedding-driven, adaptive reranking models to advance the capabilities of RAG systems and pave the way for more intelligent information retrieval applications.