HiRAG

Why we included this project

HiRAG is a lighter alternative to the heavier graph RAG stacks if your job is question answering over a large private document collection. It keeps the knowledge base in hierarchical levels and pulls from local, global, and bridge knowledge during a query, which is how it answers questions that span several documents or need a corpus-wide view. Setup is simple: install the package, load your text, and call the query method, with config files and example scripts for OpenAI, DeepSeek, and ChatGLM backends. An evaluation harness comes along too, with scripts for context extraction, graph insertion, and answer scoring across retrieval modes, so you can benchmark it against plain RAG or other graph RAG systems on your own data. If you want a research-backed, MIT-licensed starting point for hierarchical retrieval without building the indexing machinery from scratch, this.