GraphRAG Open Source Comparison 2026: Microsoft, LightRAG, KAG, HippoRAG, PathRAG

Forum topic · ✨步子哥 · 2026-06-30

Summary

An in-depth comparison of nine leading open-source GraphRAG projects evaluated on architecture, cost, query modes, incremental updates, multimodal support, and production readiness. Microsoft GraphRAG leads global summarization via Leiden community detection but incurs high LLM indexing and Map-Reduce query costs; LightRAG offers dual-level retrieval with strong incremental merging and reports up to 12x lower latency and 99% fewer API calls than GraphRAG. KAG (Ant Group) targets professional domains with schema-constrained logical form reasoning for auditable multi-hop inference. HippoRAG emulates hippocampal memory using Personalized PageRank for single-step multi-hop retrieval. PathRAG adds resource-allocation pruning on top of LightRAG to reduce noise and token usage. nano-graphrag is an 800-line educational implementation; Yuxi-Know provides a GUI platform; NebulaGraph and Neo4j serve as graph-database infrastructure. The article includes a selection decision tree, a hybrid LightRAG+GraphRAG deployment strategy, common pitfalls, and trends such as LazyGraphRAG reducing indexing cost to 0.1%, multimodal ingestion, and the emerging Agentic GraphRAG paradigm, with data current as of June 2026.

Why GraphRAG Matters

Traditional RAG slices documents into chunks and retrieves by vector similarity, returning locally similar fragments with no cross-document links. GraphRAG instead weaves text fragments into a knowledge network: entities become nodes, relations become edges, and communities become clusters. Retrieval traverses the graph rather than matching chunks. The trade-off is cost: building the graph usually requires large numbers of LLM calls.

Projects Compared

| Project | GitHub | Stars (Jun 2026) | Maintainer | Positioning | |---------|--------|------------------|------------|-------------| | Microsoft GraphRAG | microsoft/graphrag | 31k+ | Microsoft Research | Community detection + summarization, global analysis leader | | LightRAG | HKUDS/LightRAG | 29k+ | HKUDS (Hong Kong University) | Lightweight, incremental, cost-effective | | nano-graphrag | gusye1234/nano-graphrag | 2k+ | Individual | 800-line reference implementation for learning | | KAG | OpenSPG/KAG | 8k+ | Ant Group + OpenKG | Logical reasoning, professional domains | | HippoRAG | OSU-NLP-Group/HippoRAG | 3k+ | OSU NLP Group | Hippocampal-inspired memory, PPR diffusion | | PathRAG | BUPT-GAMMA/PathRAG | 1k+ | BUPT GAMMA Lab | Path pruning for noise reduction | | Yuxi-Know | xerrors/Yuxi-Know | 4k+ | Individual | GUI wrapper around LightRAG | | NebulaGraph | vesoft-inc/nebula | 12k+ | vesoft | Distributed graph database, infrastructure layer | | Neo4j RAG | neo4j/neo4j-graphrag-python | 10k+ | Neo4j | Cypher-based graph retrieval |

Key Findings

1. Microsoft GraphRAG — Heavyweight champion, heavyweight cost

Pipeline: chunking → LLM entity/relation extraction → knowledge graph → Leiden hierarchical community detection → LLM community summaries → query-time traversal of community summaries.
Query modes: Local Search (entity-centric), Global Search (community-summarization Map-Reduce), and DRIFT Search (hybrid).
Cost pain point: entity extraction per chunk plus per-community summarization yields very high API usage. On the same dataset (Chinese white-text version of the first nine chapters of *Journey to the West*), indexing cost was reported as roughly 10x that of LightRAG.
Incremental update: 0.5.0 introduced partial updates, but community structure still needs recomputation.
Latest: LazyGraphRAG claims indexing cost down to 0.1% of the original via "cluster first, summarize on demand"; not yet fully merged into the main repo.
Best for: financially capable teams needing global analysis and complex multi-hop reasoning, such as financial research, academic literature review, and industry trend scanning.

2. LightRAG — Lightweight dark horse with strong incremental updates

Pipeline: chunking → LLM entity/relation extraction → graph + Key-Value index for nodes and edges → vector DB for embeddings → dual-level retrieval.
Dual-level retrieval:
Low-level (Local): extract concrete keywords, vector-match graph nodes, fetch one-hop neighbors.
High-level (Global): extract abstract keywords, vector-match graph edges (relations), fetch both endpoints.
Hybrid (Local + Global) and Mix (Local + Global + Naive vector) modes available.
Incremental update: new documents merge via set union; no community or full-index recomputation. Ideal for frequently changing enterprise data.
Reported benchmarks (from NanGePlus/LightRAGTest):
Retrieval efficiency up 99.98% vs GraphRAG
Latency reduced ~12x
Per-query API calls reduced by >99%
Token consumption reduced >12x
Multimodal: integrates with RAG-Anything for PDF, Office, images, tables, and math formulas.
Trade-off: weaker complex reasoning than GraphRAG due to no community summary mechanism.
Best for: startups, SMEs, customer support/Q&A, frequently updated data, and cost-sensitive deployments.

3. nano-graphrag — 800 lines to understand GraphRAG

Educational/research tool, not production-grade.
Compresses core logic (entity extraction, community detection, Local/Global queries) to ~800 lines.
Pluggable LLM/embedding (default GPT-4o + text-embedding-3-small), incremental insert via md5-hash dedup, networkx + milvus-lite by default (swappable to Neo4j, FAISS, Ollama).
Limitation: each new document triggers community and report recomputation, unlike LightRAG's union merge.
Best for: learning GraphRAG internals, prompt/model experiments.

4. KAG — Ant Group's logical scalpel for professional domains

Built on OpenSPG using the LLMFriSPG framework inspired by the DIKW hierarchy.
Three core innovations:

Mutual indexing between graph nodes and text chunks

Logical form planner

Knowledge alignment

Unlike LLM-only extraction in GraphRAG, KAG adds schema constraints and structured logical forms.
Best for: medical, legal, and financial domains requiring auditable reasoning chains.

5. HippoRAG — Hippocampal-inspired oddball

Treats the LLM as neocortex for feature abstraction and the knowledge graph + Personalized PageRank (PPR) as hippocampus for indexing/retrieval.
Single-step retrieval enables complex multi-hop reasoning without iterative LLM calls.
More academic than production-ready; smaller community and fewer deployment cases.
Best for: multi-hop reasoning research and teams building on top of it.

6. PathRAG — Noise-reduction patch for LightRAG

Argument: GraphRAG and LightRAG return too much redundant information, wasting tokens and hurting answer quality.
Two key changes:
Streaming pruning algorithm: DFS-fetch three-hop neighbors, then prune low-weight branches using a resource-allocation strategy (more outgoing edges → more allocated resources → higher weight).
Path-guided prompting: describe path relations in text rather than dumping raw graph data to the LLM.
Supports Local, Global, Hybrid modes with pruning applied to each result set.
Best for: LightRAG users needing less noise and lower token usage.

7. Yuxi-Know — Suit for LightRAG

Application-layer platform: LightRAG underneath, LangChain v1 + FastAPI + Vue on top.
Full-link visualization (dashboard, knowledge-base view, graph explorer, model config), MinerU PDF parsing, user/role permissions, and ScrapeGraphAI for web scraping.
Best for: teams wanting GraphRAG without code; internal knowledge bases with user management.

8. NebulaGraph / Neo4j RAG — Infrastructure layer

Graph databases, not GraphRAG frameworks per se.
NebulaGraph: distributed storage-compute separation, trillion-scale edges/vertices, 99.999% availability, suited for TB+ deployments.
Neo4j RAG: Cypher-based graph retrieval for relationship-heavy applications.
Best for: enterprises with existing graph DB infrastructure building custom GraphRAG systems.

Selection Decision Tree

1. Data scale

TB+ / trillion edges → NebulaGraph + custom layer
Millions to billions → continue
Tens of thousands of documents → any framework works

Query type

Global summary ("what is this dataset mainly about?") → GraphRAG
Multi-hop reasoning ("relationship between A and B?") → KAG (auditable) or HippoRAG (low cost)
Local entity lookup → LightRAG (fast, cheap)
Balanced → LightRAG Hybrid or PathRAG

Update frequency

Static → GraphRAG
Daily new data → LightRAG / PathRAG
Real-time streaming → LightRAG

Budget and team

Generous budget + expert team → GraphRAG or KAG
Limited budget + small team → LightRAG
No-code → Yuxi-Know
Research/learning → nano-graphrag

Domain needs

Medical/legal (auditable) → KAG
Financial research (global view) → GraphRAG
Customer support/Q&A → LightRAG
Multi-hop research → HippoRAG
Noise-sensitive → PathRAG

Practical Hybrid Deployment

A validated hybrid strategy:

1. Route day-to-day queries to LightRAG for speed and cost — it handles ~90% of cases. 2. Route complex analysis to GraphRAG when LightRAG confidence is low or the user explicitly requests global analysis. 3. Share the underlying knowledge graph storage to avoid rebuilding indexes.

Benefit: spend GraphRAG-level cost only where it matters. Cost: operating two systems requires dedicated engineering.

Common Pitfalls

Believing GraphRAG replaces vector RAG. It does not. Vector RAG is often better for precise single-chunk lookup. Hybrid retrieval usually wins.
Underestimating indexing cost. Run a small pilot first; GraphRAG's per-chunk extraction plus per-community summarization can be an order of magnitude more expensive than expected.
Using default English prompts for Chinese data. Entity extraction quality drops without prompt tuning. KAG has native Chinese optimization.
Using nano-graphrag in production. Each insert triggers community recomputation; it is a learning tool, not a production tool.
Ignoring LazyGraphRAG. If its claimed 0.1% indexing cost is fully open-sourced, the selection landscape may shift significantly.

Trends to Watch

Cost collapsing: GraphRAG → LightRAG → LazyGraphRAG trend shows 1/12 and then ~1/1000 indexing/query cost reductions, lowering the entry barrier.
From general to domain-specific: KAG-style schema-constrained logical reasoning will outcompete general frameworks in medical, legal, and financial ToB use cases.
Multimodal fusion: LightRAG (RAG-Anything) and Yuxi-Know (MinerU) extend GraphRAG from text to all-document formats.
Agentic GraphRAG: convergent design where agents decide when to use graph retrieval, vector retrieval, or both. LangGraph + LightRAG experiments are already underway.

Quick 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-graphrag-python | — |

> Data current as of June 2026. Star counts and project status change continuously; verify against live GitHub data.