Why traditional RAG falls short
Conventional RAG splits documents into fragments and retrieves by vector similarity, so it excels at finding "the most similar chunk" but fails at cross-document reasoning, multi-hop questions, and global summarization. GraphRAG rebuilds scattered fragments into a knowledge network: entities as nodes, relations as edges, communities as clusters — retrieval walks the graph instead of only matching text.
Key points
- Microsoft GraphRAG (31k+ stars): LLM entity/relation extraction → Leiden hierarchical community detection → community summaries. Offers Local, Global, and DRIFT search. Indexing cost is extremely high (reportedly ~10x LightRAG on the same dataset); Global Search uses Map-Reduce with many LLM calls. Incremental indexing (0.5.0) still recomputes communities. Best for high-value, low-frequency global analysis (financial research, literature reviews).
- LightRAG (29k+ stars, HKU HKUDS): dual-level retrieval — vector search over nodes (Local/concrete keywords) and edges/relations (Global/abstract keywords), plus Hybrid and Mix modes. Incremental updates merge as unions without recomputing communities. Reported benchmarks vs GraphRAG: ~99.98% retrieval efficiency gain, ~12x lower latency, >99% fewer API calls, >12x fewer tokens (only two LLM calls per query). Multimodal support via RAG-Anything. Weaker at deep logical reasoning.
- nano-graphrag (2k+): ~800-line GraphRAG core for learning/research; recomputes communities on every insert — not production-ready.
- KAG (8k+, Ant Group + OpenKG): built on OpenSPG with DIKW-inspired LLMFriSPG knowledge representation; mutual indexing of knowledge and text chunks for auditable reasoning, a logical-form planner decomposing questions into operator chains, and semantic knowledge alignment. Native Chinese optimization. Ideal for healthcare, legal, finance.
- HippoRAG (3k+, OSU NLP): LLM as "neocortex," graph + Personalized PageRank as "hippocampus"; single-step PPR diffusion enables multi-hop reasoning cheaply. Academic-oriented, low production adoption.
- PathRAG (1k+, BUPT GAMMA): a LightRAG-derived denoising patch — streaming path pruning via DFS three-hop neighbors and resource-allocation-based branch pruning, plus path-guided prompts describing graph structure in text. Lower noise and token use.
- Yuxi-Know (4k+): full GUI platform (LangChain v1 + FastAPI + Vue) wrapping LightRAG; MinerU PDF parsing, permissions, dashboards, ScrapeGraphAI integration. No-code GraphRAG.
- NebulaGraph / Neo4j RAG: distributed graph database infrastructure (trillion-scale edges, 99.999% availability) and Cypher-based graph retrieval for building your own GraphRAG stack.
- GraphRAG does not replace vector RAG — for precise text retrieval, plain vector RAG is faster and more accurate; hybrid retrieval often wins.
- Indexing costs are easy to underestimate — pilot on a small dataset before full runs.
- Default prompts are English-tuned; Chinese entity extraction quality drops without prompt adjustment (KAG is the exception with native Chinese optimization).
- nano-graphrag is not production-grade.
- Watch LazyGraphRAG: Microsoft claims ~0.1% of original indexing cost via cluster-then-summarize-on-demand; not yet fully merged.
- Cost collapse: GraphRAG → LightRAG (1/12 cost) → LazyGraphRAG (three orders of magnitude lower claimed), moving GraphRAG from enterprise-only to mainstream.
- Domain specialization: schema constraints + logical reasoning (KAG) deliver more ToB value than general-purpose frameworks.
- Multimodal fusion: RAG-Anything, MinerU — "any format into the graph."
- Agentic GraphRAG: agents deciding when to use graph vs vector retrieval (e.g., LangGraph + LightRAG).
Selection decision tree
1. Data scale: TB+/trillion edges → NebulaGraph + custom layer; millions to hundreds of thousands of docs → any framework below. 2. Query type: global summarization → GraphRAG; auditable multi-hop → KAG (or HippoRAG for low cost); entity-specific lookups → LightRAG; balanced → LightRAG Hybrid or PathRAG. 3. Update frequency: static data → GraphRAG; daily/real-time updates → LightRAG/PathRAG. 4. Budget/team: large budget + expert team → GraphRAG or KAG; small team → LightRAG; no-code → Yuxi-Know; learning → nano-graphrag. 5. Domain needs: auditable chains → KAG; financial global analysis → GraphRAG; real-time Q&A → LightRAG; multi-hop research → HippoRAG; noise-sensitive → PathRAG.
A pragmatic hybrid architecture
1. Route ~90% of daily queries through LightRAG (speed + cost control). 2. Escalate low-confidence or explicitly global-analysis queries to GraphRAG. 3. Share the underlying knowledge graph storage to avoid duplicate indexing. Trade-off: running two systems adds maintenance overhead.
Common pitfalls
Trends
Project links
| Project | GitHub | Paper | |---|---|---| | Microsoft GraphRAG | github.com/microsoft/graphrag | arxiv.org/abs/2404.16130 | | LightRAG | github.com/HKUDS/LightRAG | arxiv.org/abs/2410.05779 | | nano-graphrag | github.com/gusye1234/nano-graphrag | — | | KAG | github.com/OpenSPG/KAG | arxiv.org/abs/2409.13731 | | HippoRAG | github.com/OSU-NLP-Group/HippoRAG | arxiv.org/abs/2405.14831 | | PathRAG | github.com/BUPT-GAMMA/PathRAG | — | | Yuxi-Know | github.com/xerrors/Yuxi-Know | — | | NebulaGraph | github.com/vesoft-inc/nebula | — | | Neo4j RAG | github.com/neo4j/neo4j-rag-framework | — |
> Data as of June 2026; star counts and project status change continuously — check GitHub for current numbers.