SBR-RAG: Efficient Subgraph-Based RAG with Lightweight Filter

Contemporary artificial intelligence (AI), particularly Large Language Models (LLMs), focuses on emulating human capabilities for knowledge acquisition and utilization. Relying solely on pre-training data is insufficient for up-to-date knowledge access, yet continuously retraining or fine-tuning LLMs with internal data is impractical. Many existing Retrieval-Augmented Generation (RAG) [1], [2], [3], [4] methods significantly enhance LLMs' access to external knowledge through vector-based or graph-structured retrieval from knowledge bases. However, the complexity and diversity of real-world knowledge bases hinder RAG systems from effectively retrieving complex multi-hop data and abstractly related information; moreover, increased volumes of erroneous retrieved data readily induce LLM hallucinations. To address these limitations, we propose SBR-RAG, a knowledge retrieval and filtering framework. SBR-RAG identifies relevant documents within a query-relevant subgraph and employs a lightweight discriminator for filtering, significantly improving accuracy while mitigating hallucinations. Experiments demonstrate that our method achieves superior Recall and F1 scores across multiple public datasets, demonstrating superior performance over many existing strong baselines.