LightRAG: the knowledge-graph RAG engine under RAG-Anything, Coeurdar…
A RAG engine from the HKUDS lab (MIT, paper accepted at EMNLP 2025) that maintains a knowledge graph alongside vectors. Where a plain RAG cuts the corpus into chunks and returns the ones nearest a question, this one also extracts entities and the relations between them, then queries both layers together: precise detail comes from the vectors, the wider picture from the graph. It is the layer sitting under RAG-Anything, already in the library: since May 2026 RAG-Anything has been merged into LightRAG, which therefore absorbs rich-document ingestion (images, tables, scanned PDFs). Installed through the lightrag-hku Python package.
Strengths
- Dual graph and vector layer: local detail and cross-corpus view in a single query
- Markedly lower indexing cost than GraphRAG for a benefit of the same nature
- An EMNLP 2025 paper and an identifiable lab behind it, not a demo repository
Limitations
- On a corpus with no recurring entities, the graph costs to build and returns nothing
- Relation quality depends on the extraction model, and a wrong graph beats no graph only in appearance
- A Python framework to configure, not a managed service: budget a real integration phase
Best for
- An internal corpus where the same accounts, products or decisions cross dozens of documents
- Answering cross-cutting questions rather than one-shot lookups
- A team that trialled GraphRAG and backed off at the cost of the indexing pass