LightRAG

A graph-based RAG method built for speed and cheap updates — LightRAG: Simple and Fast Retrieval-Augmented Generation (Guo, Xia, Yu, Ao, Chao Huang; HKU Data Science group / HKUDS; arXiv:2410.05779, Oct 2024). Its pitch against Microsoft GraphRAG is that you don’t need expensive community-summary map-reduce to get the benefits of a knowledge-graph: a lighter graph index plus keyword-driven retrieval gets most of the way at a fraction of the cost, and — its signature feature — it updates incrementally instead of rebuilding.

How it works

• Graph-indexed text — an LLM extracts entities and relations into a knowledge-graph, stored alongside vector embeddings of nodes/edges. Retrieval fuses the two: the graph supplies related entities and their connections, the vectors supply semantic match.
• Dual-level retrieval — every query is decomposed into low-level keywords (specific entities, concrete terms) and high-level keywords (themes, abstract topics). Low-level retrieval pulls precise facts and their neighbors; high-level retrieval pulls broad thematic context. Combining both is what lets one mechanism serve narrow and broad questions, rather than GraphRAG’s separate local/global modes.

The signature feature — incremental updates

Microsoft GraphRAG precomputes community structure and summaries over the whole corpus, so new documents mean re-running community detection and re-summarizing. LightRAG’s incremental update algorithm appends new entities/edges to the existing graph without that rebuild — “timely integration of new data” as the corpus grows. That makes it the cluster’s clearest answer to the changing-corpus cost problem (a different cut at the accumulation theme than the bi-temporal temporal-knowledge-graph, which is about which fact is current rather than cheap to add).

How it lands against the others

On its own (author-run, LLM-judged) eval it reports gains in retrieval accuracy and efficiency over NaiveRAG, RQ-RAG, HyDE, and GraphRAG. The neutral graphrag-bench places it in the middle of the cost range — ≈10⁴ tokens/query, ~4× cheaper than Microsoft GraphRAG’s global mode but ~10× more than HippoRAG2 — and, like every graph method there, task-conditional versus plain vector RAG (it helps multi-hop and thematic questions more than simple fact lookup). So within the cluster: GraphRAG optimizes for global sensemaking, HippoRAG2 for multi-hop accuracy at low cost, and LightRAG for simplicity + cheap incremental updates. Open-source (hkuds/LightRAG), senior author chao-huang — both now paged, closing the “authors/HKUDS noted inline (thin)” caveat this page carried from 2026-06-29.

What was built on it (added 2026-08-04)

rag-anything extends LightRAG to multimodal documents — images, tables and LaTeX equations promoted to graph entities, with a MinerU parsing stage in front. Same group, same senior author, and it keeps the incremental-update property above rather than reintroducing the rebuild cost. So this page is now the base of a two-system lineage, and the corpus’s evidence about LightRAG’s cost position partly transfers: a multimodal extension of a mid-cost graph method starts from mid cost.

Related

rag-anything · hkuds · chao-huang · graphrag-bench · graphrag · from-local-to-global-graphrag · hipporag2 · raptor · leanrag · knowledge-graph · retrieval-augmented-generation · temporal-knowledge-graph