LightRAG is a retrieval-augmented generation framework presented at EMNLP 2025 that emphasizes simplicity and speed in knowledge graph-based RAG systems. The repository, maintained by the HKUDS team, implements a dual-level retrieval approach combining vector-based and graph-based methods to enhance both the quality and efficiency of information retrieval for large language models.
The framework is written in Python and distributed via PyPI as the lightrag-hku package.
LightRAG has been classified by GitGenius across multiple domains including RAG Framework, LLM Applications, Generative AI, Information Retrieval, Natural Language Processing, AI Development, Modular Design, Evaluation, Deployment, and Data Processing.
The framework supports multiple storage backends and deployment options. Users can deploy LightRAG via PyPI installation, source installation, or Docker Compose. The repository includes an interactive setup wizard that generates configured environment files and Docker Compose configurations, facilitating both local and cloud deployments. For offline or air-gapped environments, the project provides a dedicated offline deployment guide.
Recent feature additions demonstrate the project's evolution toward comprehensive RAG capabilities. The system now supports multimodal content processing through integration with MinerU and Docling services, four selectable text chunking strategies (Fix, Recursive, Vector, and Paragraph), and role-specific LLM configuration with four distinct roles: EXTRACT, QUERY, KEYWORDS, and VLM. OpenSearch integration provides unified storage backend support, while reranker functionality has been added to boost performance for mixed queries. The framework includes document deletion with automatic knowledge graph regeneration, citation functionality for source attribution, and integration with RAGAS for evaluation and Langfuse for tracing.
Storage flexibility is a key feature, with support for Neo4j, MongoDB, PostgreSQL, and OpenSearch as backend options. The project includes a WebUI for intuitive knowledge graph insertion, querying, and visualization. The framework also supports local deployment of embedding, reranking, and storage backends via Docker, enabling fully self-contained deployments.
The repository maintains active community channels including a Discord server and WeChat group, with documentation available in English, Chinese, and Japanese. The project has spawned related initiatives including RAG-Anything for multimodal RAG, VideoRAG for long-context video understanding, and MiniRAG for simplified RAG with smaller models, indicating a growing ecosystem around the core LightRAG framework.
