Developers building RAG applications who want graph structured context alongside standard vector search.

LightRAG is a research backed retrieval augmented generation framework that builds and queries a knowledge graph alongside vector embeddings, supporting local, global, hybrid, and naive retrieval modes.

Self hosted as a Python library or server, pluggable into PostgreSQL, Neo4j, MongoDB, Milvus, and other storage backends.

Published as an EMNLP 2025 paper and actively developed on GitHub since late 2024.

Outperforms naive and prior graph based RAG methods like GraphRAG on multi domain benchmarks while remaining lightweight to run.

Requires an LLM and embedding model of the user's choice plus a supported graph or vector storage backend.

Open source under MIT on GitHub, no account required, maintained by the University of Hong Kong Data Science Lab, HKUDS.