LightRAG is a graph-enhanced retrieval-augmented generation (RAG) framework from HKUDS (paper: arXiv:2410.05779, EMNLP 2025; GitHub: HKUDS/LightRAG, 35.6k stars). Its stated goal: keep the reasoning capabilities of graph-based RAG while cutting costs to roughly 1/50 of Microsoft GraphRAG.
The Cost Problem with GraphRAG
Microsoft GraphRAG replaces flat text chunks with knowledge graphs, enabling cross-document global reasoning—but at enormous cost. On the Legal dataset:
- Retrieval: ~610,000 tokens (610 communities × 1,000 tokens each) and hundreds of API calls per query
- Indexing: ~360,000 tokens
- Incremental updates: rebuilding all 1,399 communities costs ~14M tokens
- Every entity/relationship is a key-value pair
- Key: entity name or relationship description (used for vector retrieval)
- Value: a text summary of that entity/relationship's context in the source documents
- Low-level: targets specific entities/relations (e.g., "Who wrote Pride and Prejudice?" → the Jane Austen node). Suited to factual queries.
- High-level: targets abstract themes (e.g., "How does AI influence modern education?" → multiple relation paths around AI and education). Suited to synthesis queries.
- Hybrid: both levels' results are concatenated for the LLM. Ablations show removing either level significantly hurts performance.
- Graph-enhanced systems decisively beat pure vector baselines, especially on large datasets (Legal)
- LightRAG outperforms GraphRAG on Agriculture, CS, and Legal, with advantages growing with data scale
- LightRAG consistently leads on diversity, thanks to dual-level retrieval covering both facts and themes
- Guo Z, Xia L, Yu Y, et al. LightRAG: Simple and Fast Retrieval-Augmented Generation. *EMNLP 2025*. arXiv:2410.05779
- Edge D, et al. From Local to Global: A Graph RAG Approach to Query-Focused Summarization. *Microsoft Research*. arXiv:2404.16130
- GitHub: https://github.com/HKUDS/LightRAG
The root cause is GraphRAG's reliance on pre-computed community summaries via LLM calls.
LightRAG's Core Architecture
1. Graph-Based Key-Value Indexing (No Community Summaries)
LightRAG chunks documents and extracts entities and relationships with an LLM, but skips community summarization entirely. Instead:
Retrieval matches query keywords against keys and pulls the corresponding values—replacing pre-computed global summaries with on-demand entity descriptions. A deduplication function merges same-named entities across documents.
2. Dual-Level Retrieval
3. Graph + Vector Hybrid Retrieval Pipeline
1. LLM extracts local keywords (entities) and global keywords (themes) from the query 2. Vector matching: local keywords → entity nodes; global keywords → relation edges 3. First-hop neighbors are fetched to build subgraph context 4. All value texts are concatenated with the query for answer generation
Entire retrieval takes one API call with <100 input tokens.
Experimental Results (UltraDomain Benchmark)
UltraDomain covers Agriculture, CS, Legal, and Mix datasets (600K–5M tokens each), evaluated via LLM-as-a-Judge (GPT-4o-mini) on comprehensiveness, diversity, empowerment, and overall. Baselines: NaiveRAG, RQ-RAG, HyDE, GraphRAG.
| Metric | NaiveRAG | RQ-RAG | HyDE | GraphRAG | LightRAG | |---|---|---|---|---|---| | Agriculture Overall | ~30% | ~35% | ~40% | ~45% | ~60% | | CS Overall | ~25% | ~30% | ~35% | ~40% | ~55% | | Legal Overall | ~20% | ~22% | ~25% | ~35% | ~55% | | Mix Overall | ~30% | ~35% | ~40% | ~50% | ~55% |
*(Win-rate trends per the paper; exact figures in Table 1.)*
Key findings:
Ablations
| Variant | Change | Result | |---|---|---| | -High | Remove high-level retrieval | Broad performance drop; fails on synthesis queries | | -Low | Remove low-level retrieval | Breadth retained but depth lost; weak on factual queries | | -Origin | Drop raw text, graph values only | No significant drop; sometimes improves—graph indexing filters source noise |
Cost Comparison
Indexing (Legal): GraphRAG ~360,000 tokens vs. LightRAG ~29,000 tokens (≈1/12).
Retrieval: GraphRAG 610,000 tokens / hundreds of calls vs. LightRAG <100 tokens / 1 call. Estimated per-query cost at GPT-4o-mini pricing: ~$0.30 vs. ~$0.0001—a ~3,000× gap.
Incremental updates: GraphRAG effectively re-runs everything (~14M tokens); LightRAG simply unions new entities/relations into the existing graph, with cost proportional to new documents.
Ecosystem and Limitations
Released October 2024, LightRAG has reached 35.6k GitHub stars. It is pure Python, defaults to GPT-4o-mini (any OpenAI-compatible API works), uses a nano vector store by default, and has community backends for Neo4j, PostgreSQL, MongoDB, and more.
Open questions:
1. Entity extraction quality caps retrieval quality—LightRAG lacks the redundancy of GraphRAG's community-level summaries 2. Graph growth under frequent incremental updates; long-term latency impact unreported 3. Text-only in the original release (RAG-Anything from the same team addresses multimodality) 4. Retrieval and generation remain decoupled; end-to-end alternatives are unproven
Conclusion
LightRAG's real contribution is proving that graph-enhanced RAG can be both good and cheap. By abandoning community summaries in favor of a key-value graph index and dual-level retrieval, it compresses GraphRAG's two-orders-of-magnitude cost into production-acceptable ranges—making knowledge-graph RAG viable for cost-sensitive, large-scale deployments.