HiRAG
HiRAG is an open-source Retrieval-Augmented Generation framework that leverages hierarchical knowledge structures to enhance large language model performance. Introduced in the EMNLP 2025 Findings paper, HiRAG addresses the limitations of naive RAG by organizing information into global and local graphs with bridge connections. This hierarchical approach allows the system to efficiently retrieve relevant context at different abstraction levels without requiring costly re-indexing of the entire knowledge base. The software supports integration with various LLMs including DeepSeek, ChatGLM, and OpenAI models, and is compatible with popular datasets like UltraDomain. Key features include customizable indexing workflows, hierarchical and naive retrieval modes, and comprehensive evaluation scripts for benchmarking accuracy. Users can部署 the framework locally using Python and Pip, configure parameters via YAML files, and run queries through a simple Python API or command-line interface. HiRAG is designed for research