HKUDS/LightRAG

LightRAG rewires how retrieval-augmented generation indexes and queries knowledge by building a dual-level graph structure — entities and relationships get indexed alongside raw text chunks, so queries that span multiple concepts actually find the connective tissue instead of returning five loosely related paragraphs. The differentiator over naive RAG pipelines is that retrieval is graph-aware: asking 'how does X relate to Y in this corpus' actually works, rather than returning whichever chunk mentioned both words most recently. It supports incremental updates to the graph without full re-indexing, which is the thing that makes it usable in production rather than just in demos. The score here is a genuine signal — this one has traction in the research and engineering community simultaneously. Reservation: setup assumes you are comfortable wiring your own LLM backend and vector store. Not a plug-and-play SaaS drop-in. -> Best for: AI engineer building knowledge-intensive RAG pipelines on proprietary document corpora