hhy-huang/HiRAG

Organizes retrieved documents into local, global, and bridge knowledge layers within a graph database, enabling multi-level contextual retrieval that significantly outperforms flat naive RAG and graph-based approaches across comprehensiveness, empowerment, and diversity metrics. Supports LLM-agnostic integration through pluggable APIs (DeepSeek, ChatGLM, OpenAI) and batch async embedding with configurable caching, while providing modular retrieval modes (hierarchical, local-only, global-only, bridge-only) for controlled knowledge access. Includes evaluation pipelines on the UltraDomain benchmark across specialized domains (Mix, CS, Legal, Agriculture) with LLM-based assessment scoring.

522 stars.