Most RAG systems flatten documents into isolated chunks, so a question that spans several entities gets fragmented context and loses the relationships that make the answer coherent. LightRAG's bet is that a knowledge graph layered over vector retrieval recovers those connections cheaply enough to run alongside, not instead of, normal embedding search.
Key Findings
- Dual-level retrieval pulls both fine-grained entity detail (low-level) and broad thematic context (high-level) in one pass, which the paper credits for stronger diversity than chunk-only retrieval.
- Incremental ingestion builds a local graph per new document and merges it via set operations, so updating the index does not mean re-indexing the corpus — a direct answer to GraphRAG's rebuild cost.
- Reported wins are uneven and honest about it: a clear 60-84.8% edge over naive RAG, but roughly a coin-flip (49.6-54.8%) against Microsoft GraphRAG, where the draw is lower overhead rather than higher accuracy.
Who It Fits
Great fit if you have a knowledge base that changes often and you want graph-aware retrieval without GraphRAG's reconstruction tax, or if you need pluggable storage backends (Postgres, Neo4j, Milvus, Qdrant) and multiple query modes. Look elsewhere if naive vector RAG already answers your queries well — the graph construction adds LLM calls and operational moving parts that only pay off when cross-entity reasoning matters.
