LazyGraphRAG vs GraphRAG: indexing and query costs

LazyGraphRAG shifts model work from indexing to answering questions. The speaker explains how noun phrases become nodes in a co-occurrence graph, while graph statistics build a community hierarchy without language-model calls. Summaries wait until a query identifies relevant material. This reduces upfront model spending, but the graph lacks typed entities, canonicalized names and explicit relationship labels.

At query time, best-first community selection combines with breadth-first exploration and iterative deepening. A model tests sentences for relevance, and the system summarizes the evidence it gathers. The relevance-test budget controls how much material it examines; the reported configurations also differ in model choice.

The module compares this approach with eager Microsoft GraphRAG and LinearRAG's token-free indexing and retrieval, using query volume, corpus volatility and question breadth to discuss costs. The speaker attributes the reported savings to Microsoft's evaluation of 5,590 news articles and 100 synthetic queries, with an LLM judging answers. Those results are not independent validation across datasets.

Availability limits matter for self-hosted use. The speaker says LazyGraphRAG was absent from Microsoft's open source GraphRAG library, cites a January 2026 release-priority statement, and identifies Microsoft Discovery on Azure as its productized implementation. GitHub recreations are described as unofficial implementations.