GraphRAG
GraphRAG uses an LLM-built knowledge graph as the retrieval index instead of (or alongside) flat text chunks. It targets questions plain vector RAG handles badly, such as “what are the main themes across this corpus?” See retrieval-augmented-generation-overview for the baseline.
Microsoft GraphRAG
The paper “From Local to Global: A Graph RAG Approach to Query-Focused Summarization” (Edge et al., arXiv 2404.16130, April 2024) uses two LLM stages: extract an entity knowledge graph, then pre-generate summaries for entity communities. At query time each community summary yields a partial answer that is merged into a final one. On corpora of about 1M tokens it beat conventional RAG on comprehensiveness and diversity for global questions. The MIT-licensed repo offers local search (specific entities), global search, DRIFT search, and LazyGraphRAG/FastGraphRAG variants. Important current status: the README says it is in maintenance mode, no longer accepting new feature PRs, only bug fixes and dependency updates, and warns that indexing can be expensive.
LightRAG
LightRAG (HKUDS, MIT, EMNLP 2025, arXiv 2410.05779) is a lighter graph-plus-vector framework with dual-level retrieval (specific facts and abstract concepts). It runs in memory by default and supports PostgreSQL, MongoDB, OpenSearch, Neo4j, Milvus, Qdrant and Memgraph; about 40k GitHub stars, with multimodal parsing (MinerU/Docling) added in 2026.
Neo4j
Neo4j is the most common graph store. Its official neo4j-graphrag Python package provides a knowledge-graph builder, pipelines, vector index management and several retrievers; it replaces the deprecated neo4j-genai.
When to use
Worth the indexing cost for corpus-wide synthesis, multi-hop entity questions and explainable provenance. For point lookups, hybrid search plus reranking is cheaper (hybrid-search-and-rank-fusion). Temporal graphs for agent memory: agent-memory-systems.
Sources
- https://arxiv.org/abs/2404.16130 (accessed 2026-09-30)
- https://github.com/microsoft/graphrag (accessed 2026-09-30)
- https://github.com/HKUDS/LightRAG (accessed 2026-09-30)
- https://neo4j.com/docs/neo4j-graphrag-python/current/ (accessed 2026-09-30)