AI Aulis

AI NEWS IN ENGLISH · 18 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

32 stories · next issue tomorrow morning · How the paper is made →

← Issue 18 November 2026Paper

In Microsoft’s test, GraphRAG Zero assembles answer context in under a second on average

AI-generated contentWritten by AI

msft.it, x.com · 2 sources · Sources published: 17 November 2026

Microsoft is bringing its graph-based GraphRAG Zero retrieval to Foundry IQ and GitHub’s Repo Mind Light. In the company’s own test, the retrieval assembled the answer context for the language model in 0.93 seconds on average.

Microsoft Research announced GraphRAG Zero in a blog post on 17 November. It is the latest generation of the company’s graph-based retrieval method for questions about an entire dataset, such as the key risks across a collection of reports. Standard RAG is poorly suited to such questions, Microsoft said. It retrieves the text chunks most similar to the question as the basis for the language model’s answer.

The original GraphRAG, released in 2024, used a large language model to extract the entities mentioned in the data and their relationships, and summarised them by topic while building the index. This required a lot of language model work. Its successor, LazyGraphRAG, moved that work to query time, so assembling the answer context could take a minute or longer.

GraphRAG Zero builds its graph from embeddings, numerical vectors that represent the meaning of text. Building the index requires no language model calls. In the real-time configuration, the language model writes only the final answer.

Microsoft tested the method on podcast transcripts and news articles, using 200 AI-generated questions: half multi-hop and half about the whole dataset. GraphRAG Zero and vector search could each retrieve 50 text chunks, and the same model wrote the answers. For whole-dataset questions, GraphRAG Zero’s answers covered 57.4 per cent of the assertions derived from the data, against 53.4 per cent for vector search. Assembling the answer context took 0.93 seconds on average, excluding answer generation.

A language model judged a larger share of the vector search chunks to be relevant. GraphRAG Zero, however, covered more of the topic clusters relevant to the question. The more representative context supported more comprehensive answers, Microsoft said, so the share of relevant chunks alone is not a sufficient measure of retrieval. The 4-percentage-point gap was measured on the company’s own questions. The figures are not measurements of the Foundry IQ service, the company said. The questions and assertions were generated with Microsoft’s open BenchmarkQED tool, which developers can also use to evaluate their own retrieval systems.

GraphRAG Zero is available as an option in Foundry IQ (Azure AI Search). GitHub Next’s Repo Mind Light uses it to retrieve background, such as architecture and past decisions, from a repository’s issues and pull requests for people and coding agents. Repo Mind Light is available as a public container that works with GitHub Agentic Workflows. Microsoft Discovery has committed to adopting the method in scientific research. The blog does not disclose pricing.

Report an error in this story