From Local to Global: A Graph RAG Approach to Query-Focused Summarization

The canonical GraphRAG paper (Edge, Trinh, Cheng, Larson et al., Microsoft Research, April 2024; arXiv:2404.16130). It introduces GraphRAG and is the primary source the graphrag page had flagged as missing — prior grounding was the T3 implementation graphrag-rs and the T2 competitor leanrag. The framing the wiki’s “fifth gap” borrows comes from here: the title’s local to global is exactly the local-vs-global query split.

The problem it names

Plain RAG retrieves a few similar chunks, which answers local questions but not global ones — “what are the main themes across the whole corpus?” — because no chunk holds the answer. Classic query-focused summarization (QFS) could answer such questions but doesn’t scale to RAG-sized collections. GraphRAG bridges the two.

The method

• Build an entity knowledge-graph — an LLM extracts entities and relationships from source chunks.
• Detect communities with Leiden — chosen for recovering hierarchical community structure efficiently. This yields levels C0 (root, fewest communities) → C3 (leaf, finest-grained).
• Pre-summarize each community into a report, at every level.
• Answer by map-reduce over summaries — prepare: shuffle and chunk the relevant community summaries; map: each chunk yields a partial answer with a helpfulness score (0–100), zero-scored answers dropped; reduce: sort by score, pack context to capacity, synthesize the global answer.

What the evaluation showed

Two ~1M-token corpora — podcast transcripts (~1.67M tokens) and news articles (~1.7M tokens). Baselines: SS (naive vector RAG) and TS (map-reduce summarization run directly over source text). LLM-judged head-to-head on comprehensiveness, diversity, empowerment, with directness as a validity control.

• Every GraphRAG level (C0–C3) beat naive vector RAG — comprehensiveness win rates 72–83% (podcast) and 72–80% (news), with comparable diversity gains.
• Cost is the headline practical finding. Root-level summaries (C0) need 9×–43× fewer tokens per query than source-text summarization (TS) and >97% fewer than leaf-level (C3) summaries — so a coarse community level buys most of the global-sensemaking quality at a fraction of the cost.

How it lands in the wiki

This resolves the graphrag sourcing gap: the method is now grounded in its defining primary source, not only in an implementation (graphrag-rs) and a competitor (leanrag). It also supplies the cost/quality knob the implementation pages couldn’t — the hierarchical levels and the C0 token-cost win — and it backs LeanRAG‘s “semantic islands” critique with the original design it critiques (GraphRAG’s community summaries are exactly the per-cluster reports LeanRAG argues are left unconnected). Microsoft’s open-source Python implementation was announced here (aka.ms/graphrag).

Tier note (T1, with caveat). The defining primary source from a major lab. But like leanrag, its comparative results are LLM-judged on the authors’ own setup — strong on method, motivated on the win-rate margins.

Related

graphrag · graphrag-rs · leanrag · knowledge-graph · retrieval-augmented-generation · leiden-algorithm · microsoft · rag-original-paper