Source: Microsoft Research: LazyGraphRAG

Microsoft Research blog on LazyGraphRAG, which defers LLM work to query time and indexes at about vector-RAG cost.

Contents

Summary

Presents a GraphRAG variant that builds its concept graph with cheap NLP noun-phrase extraction instead of LLM summarization, and spends LLM calls only at query time under a tunable relevance-test budget.

Key claims

  • Indexing cost equals vector RAG and is about 0.1% of full GraphRAG.
  • Query: best-first (vector similarity) plus breadth-first community search with iterative deepening and an LLM relevance assessor.
  • Matches GraphRAG global search quality on global queries at over 700x lower query cost; at 4% of its query cost beats all compared methods on local and global queries.
  • Evaluated on 5,590 AP news articles with 100 synthetic queries, judged by an LLM.
  • A single relevance-test budget parameter trades cost for quality.

Feeds

GraphRAG

Suggest changes →