LightRAG

Combining a knowledge graph with vector embeddings, LightRAG answers queries through dual-level retrieval: a local mode that follows entity relationships for specific questions and a global mode that reasons over higher-level themes, with hybrid, naive, and mix strategies rounding out five query modes. The framework from HKU's data intelligence lab positions itself as a lighter, cheaper alternative to Microsoft's GraphRAG, and the approach is documented in an EMNLP 2025 paper. Documents index incrementally without rebuilding the graph, parse through MinerU, Docling, or a native pipeline, and persist across storage backends including PostgreSQL, Neo4j, MongoDB, Milvus, Qdrant, and OpenSearch, with multimodal support for tables, formulas, and images. The lightrag-hku package installs from PyPI on Python 3.10 or newer, and the repository ships a REST API server with a web UI plus Docker Compose deployment. It pairs with OpenAI-compatible APIs or local models through Ollama, carries an MIT license, and has passed 38,000 GitHub stars, making it one of the most adopted open RAG frameworks.