Duolin Sun Dan Yang Yue Shen Yihan Jiao Zhehao Tan Jie Feng Lianzhen Zhong Jian Wang Peng Wei Jinjie Gu

Abstract

The Retrieval-Augmented Generation (RAG) approach enhances question-answeringsystems and dialogue generation tasks by integrating information retrieval (IR)technologies with large language models (LLMs). This strategy, which retrievesinformation from external knowledge bases to bolster the response capabilitiesof generative models, has achieved certain successes. However, current RAGmethods still face numerous challenges when dealing with multi-hop queries. Forinstance, some approaches overly rely on iterative retrieval, wasting too manyretrieval steps on compound queries. Additionally, using the original complexquery for retrieval may fail to capture content relevant to specificsub-queries, resulting in noisy retrieved content. If the noise is not managed,it can lead to the problem of noise accumulation. To address these issues, weintroduce HANRAG, a novel heuristic-based framework designed to efficientlytackle problems of varying complexity. Driven by a powerful revelator, HANRAGroutes queries, decomposes them into sub-queries, and filters noise fromretrieved documents. This enhances the system's adaptability and noiseresistance, making it highly capable of handling diverse queries. We comparethe proposed framework against other leading industry methods across variousbenchmarks. The results demonstrate that our framework obtains superiorperformance in both single-hop and multi-hop question-answering tasks.