Source: Microsoft Research, The evolution of GraphRAG: advancing systems and evaluation together (2026-11-17), announced on LinkedIn. GraphRAG Zero ships as an option in Foundry IQ (Azure AI Search) and inside GitHub Next's Repo Mind Light. The research numbers below come from Microsoft's own study and, as the post says, are not measurements of the hosted service.
Three generations
GraphRAG (2024): an LLM extracts entities and relationships, the graph is split into communities, and an LLM writes a report for each. Local questions use entity detail; global questions ("what are the main risks across these reports?") synthesize community reports. Powerful, but indexing is LLM-heavy. The paper has about 3.9k citations and the open-source library over 36k GitHub stars.
Fast indexing and LazyGraphRAG: first a concept co-occurrence graph built with conventional NLP; then LazyGraphRAG defers all generative work to query time, cutting indexing cost to about 0.1% of full GraphRAG and keeping source text instead of lossy summaries. The cost moves to query time, which can take a minute or more.
GraphRAG Zero: builds the graph from the geometry of text embeddings, so it makes no LLM calls at index time and, in its real-time configuration, none at query time before answer generation. Context is assembled in under a second, which makes the index cheap enough to rebuild continuously.
What the evaluation adds
The story is as much about measurement as retrieval. BenchmarkQED, Microsoft's open-source suite, generates questions (AutoQ) from a dataset or from representative user activities, at local or global scope, and now also generates source-grounded assertions: testable statements an answer should contain. Retrieval is scored separately from writing quality: chunk precision from LLM relevance judgments, and a cluster recall that measures topical breadth rather than document recall.
In the GraphRAG Zero study (podcast transcripts and news, 200 synthetic questions split between multi-hop and global, 50-chunk budget, same embedder and answer model): global answer assertion coverage rose from 53.4% with vector RAG to 57.4%, with context prepared in 0.93 seconds on average. GraphRAG Zero had higher cluster recall and topical fidelity; vector RAG kept higher chunk precision. The lesson the authors draw: precision alone would have steered development the wrong way, because more representative evidence produced more complete global answers.
Evaluation for agent loops
For agents whose knowledge changes as they work, the team scores graded evidence coverage: an LLM judges each chunk's support for each assertion as full, partial or none; the best support per assertion is averaged. Repeated evidence earns nothing; covering a missing fact does. With datasets, questions, assertions and judging held fixed, assertion-chunk judgments can be cached and reused across system versions, so each iteration only pays to judge new evidence. Microsoft frames this as part of loop engineering: improve, test against a fixed standard, repeat.
A related GitHub Next project, WikiKB, keeps repository knowledge as reviewable Markdown with provenance and Git history, indexed and retrieved locally.
Tradeoffs
Breadth versus precision: Zero trades some chunk precision for topical coverage; good for global and multi-hop questions, not obviously better for pinpoint factual lookup, where vector retrieval stays strong.
Embedding geometry as structure: no LLM-written entity graph means no human-readable entity and relationship descriptions; you lose the analytic artifacts of the original.
Modest headline gain: four points of assertion coverage on synthetic questions; worthwhile, but test on your own corpus.
Judge dependence: every metric relies on LLM judgments, so judge drift or bias moves all the numbers together.
Failure modes to plan for
Global questions answered from a representative but incomplete slice, with confident synthesis over missing topics.
Stale context when the index is not actually rebuilt as often as the data changes.
Evaluation sets that drift away from real user questions, because AutoQ generates synthetic ones.
Cached judgments reused after the judging prompt or model changed, silently mixing standards.
Review checklist
What share of your real questions are global or multi-hop versus pinpoint lookups?
Do you score retrieval separately from answer quality, with both precision and breadth?
Are assertion sets grounded in your sources and reviewed by a person?
Is the judge model and prompt pinned, and are cached judgments invalidated when it changes?
How often is the index rebuilt relative to how fast the corpus changes?
Related entries
R106 txtai: an all-in-one embeddings database.
R083 15 building blocks of agent memory.
R123 Proactive agents: memory that stays useful as it grows.
R107 Ten evaluation papers of 2026.