LightRAG: A Low-Cost Graph-Enhanced RAG Alternative to GraphRAG

Paper: arXiv:2410.05779 (EMNLP 2025) | Code: github.com/HKUDS/LightRAG (~35.6k stars) Authors: Zirui Guo, Lianghao Xia, Yanhua Yu, Tu Ao, Chao Huang (BUPT + HKU)

---

Key points

1. The cost problem with GraphRAG

Microsoft's GraphRAG applies knowledge graphs over documents, but its pipeline is prohibitively expensive:

• Indexing: ~360,000 tokens for community summaries
• Retrieval: ~610,000 tokens per query on the Legal dataset (610 communities × 1,000 tokens) plus hundreds of API calls
• Incremental update: ~14,000,000 tokens to rebuild all 1,399 communities

For small teams, API bills can easily exceed thousands of USD per month.

2. Core design — key-value graph + dual-level retrieval

#### 2.1 Graph-based text indexing LightRAG extracts entities and relations from each chunk via an LLM but skips community summarization entirely. The resulting index is a flat key-value store:

• Key: entity name / relation description (used for vector matching)
• Value: a short text description summarizing that entity or relation's context in the source documents

A deduplication function merges same-named entities across documents, keeping the graph bounded. Retrieval becomes simple key lookup followed by value retrieval — replacing GraphRAG's "precompute all community reports" with "fetch entity descriptions on demand."

#### 2.2 Dual-level retrieval

• Low-level: precise matches on named entities and relations (e.g., "Who wrote Pride and Prejudice?" → Jane Austen node + 1-hop neighbors). Best for factual queries.
• High-level: matches broader themes by aggregating multiple relation paths (e.g., "How does AI influence modern education?"). Best for synthesis queries.
• Both levels are concatenated and fed to the LLM. Ablations show removing either level degrades performance significantly.

#### 2.3 Hybrid graph + vector retrieval flow 1. LLM extracts local keywords (entities) and global keywords (themes) from the query. 2. Vector search retrieves matching entity nodes (local) and relation edges (global). 3. One-hop neighbors are added to build a subgraph context. 4. All value texts are concatenated with the query for answer generation.

The entire retrieval phase needs a single API call with under 100 input tokens — two orders of magnitude cheaper than GraphRAG.

3. Experimental results — UltraDomain benchmark

Setup: Agriculture, CS, Legal, Mix domains (600k–5M tokens each), evaluated with GPT-4o-mini as judge across four dimensions: Comprehensiveness, Diversity, Empowerment, Overall. Baselines: NaiveRAG, RQ-RAG, HyDE, GraphRAG.

Headline findings (overall win-rate, approximate trends):

| Domain | NaiveRAG | RQ-RAG | HyDE | GraphRAG | LightRAG | |---|---|---|---|---|---| | Agriculture | ~30% | ~35% | ~40% | ~45% | ~60% | | CS | ~25% | ~30% | ~35% | ~40% | ~55% | | Legal | ~20% | ~22% | ~25% | ~35% | ~55% | | Mix | ~30% | ~35% | ~40% | ~50% | ~55% |

Key takeaways:

• Pure-vector RAG collapses on complex domains (NaiveRAG/HyDE win rates vs. LightRAG drop to ~20% on Legal).
• LightRAG consistently beats GraphRAG; advantage grows with dataset size.
• Diversity is a standout strength, thanks to dual-level coverage of facts and themes.

Ablation results:

| Variant | Change | Effect | |---|---|---| | -High | Remove high-level retrieval | Performance drops overall; over-focuses on local detail, fails synthesis queries | | -Low | Remove low-level retrieval | Broader but shallower; poor on factual queries | | -Origin | Drop raw chunks, keep only graph values | No significant loss — sometimes improves — confirming the graph filters original-text noise |

The third result is notable: LightRAG can operate without storing original chunks, further reducing storage cost.

4. Cost comparison (Legal dataset, GPT-4o-mini pricing)

#### Indexing

| System | Tokens | API calls | |---|---|---| | GraphRAG | ~360,000 | Multiple | | LightRAG | ~29,000 | Proportional to chunk count |

Ratio: ~1/12 of GraphRAG.

#### Retrieval

| System | Tokens | API calls | |---|---|---| | GraphRAG | 610,000 | Hundreds | | LightRAG | <100 | 1 |

Estimated per-query cost: GraphRAG ~$0.30 vs. LightRAG ~$0.0001 — roughly a 3,000× reduction.

#### Incremental update

| System | Approach | Tokens | |---|---|---| | GraphRAG | Rebuild communities + regenerate summaries | ~14,000,000 | | LightRAG | Union new graph into existing graph | Proportional to added docs |

LightRAG performs true incremental updates: new entities and relations are merged in, leaving existing links intact.

5. Open-source ecosystem

Released October 2024, LightRAG reached ~35.6k stars within a year. Features:

• Pure Python with minimal dependencies (LLM + vector store + graph DB)
• Defaults to GPT-4o-mini; any OpenAI-compatible API works
• Ships with a built-in nano vector store
• Community backends: Neo4j, PostgreSQL, MongoDB, and others

Its adoption stems not only from solid research but also from being affordable to run — free-tier API usage is feasible, unlike with GraphRAG.

6. Limitations and open questions

1. Entity extraction is the ceiling. Missed entities/relations cannot be recovered at retrieval time. GraphRAG's community summaries serve as a soft redundancy that LightRAG deliberately drops. 2. Graph growth under continuous updates. The paper does not report long-term retrieval latency after many incremental additions — a key deployment concern. 3. Text-only. Non-text modalities (tables, figures, images) are not natively supported; the same team's later RAG-Anything addresses this. 4. Pipeline vs. end-to-end training. Whether VLM-style or jointly trained models can match this efficiency remains an open research question.

7. Takeaway

LightRAG's central insight is not that graph-enhanced RAG is better — GraphRAG already proved that. The insight is that graph-enhanced RAG can be both effective and cheap. By replacing community summarization with key-value indexing and adopting dual-level retrieval, LightRAG compresses GraphRAG's two-orders-of-magnitude cost into a production-viable range, turning graph RAG from a lab demo into deployable infrastructure. For cost-sensitive, large-scale deployments, it is currently one of the strongest options available.

---

References

• 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.* arXiv:2404.16130
• Qian J, et al. *UltraDomain: A Benchmark for LLM-based RAG Systems.* 2024
• GitHub: https://github.com/HKUDS/LightRAG