Common Core Ontologies
Home/Blog/Microsoft Foundry IQ adds knowledge graph entities and topics for AI agents
NEWS

Microsoft Foundry IQ adds knowledge graph entities and topics for AI agents

Microsoft Foundry IQ adds knowledge graph entities and topics for AI agents

At Microsoft Ignite 2026 on November 17, Microsoft put knowledge graphs into public preview inside Foundry IQ (formerly Azure AI Search), giving enterprise AI agents two new ways to read indexed content: an entity graph that connects facts across documents, and a topic hierarchy that organizes material by subject for synthesis across sources. Redmond magazine's coverage of the Ignite announcements frames the move alongside a new specialized retrieval model and generally available, permission-aware SharePoint indexing as part of a broader push to make Foundry IQ's retrieval layer structure-aware rather than purely vector-based. For teams planning data engineering services around agentic retrieval, the release is a signal that major platform vendors are converging on graph-plus-ontology grounding as the default, not an advanced add-on.

What ships in the preview

Knowledge Graph Entities extracts structured facts and relationships from documents already indexed by Foundry IQ, so an agent can answer a query that spans multiple sources (who reports to whom, which contract supersedes which, which incident ticket is linked to which outage) without the developer hand-building that logic in application code. Organizations can attach their own ontology to a knowledge source, or bring an existing graph and query it directly in Graph Query Language (GQL) alongside natural-language queries through the standard knowledge base interface. Knowledge Graph Topics does a complementary job: it clusters documents into a browsable topic hierarchy so retrieval can pull evidence from across a subject area instead of a single best-matching chunk.

The numbers Microsoft is citing

Microsoft's own benchmark, published in the Azure AI Foundry blog post announcing the feature, reports that entity-graph retrieval answered 93.5% of structured queries correctly against 26.1% for retrieval-augmented generation (RAG) alone, while using 64.3% fewer tokens per query. Topic-based retrieval is credited with lifting evidence recall by up to 13.2 percentage points on its own, and the combined entity-plus-topic approach recovered 92% of required facts using 52% fewer tokens than RAG alone. Microsoft frames the graph as a complement to RAG rather than a replacement, recommending teams layer retrieval strategies based on whether a given query needs connected facts, thematic breadth, or straightforward passage similarity.

Why the framing matters beyond Azure

The specific numbers are Microsoft's own benchmark and have not been independently reproduced, so they should be read as a vendor claim rather than a settled result. What is harder to dispute is the design choice underneath them: Foundry IQ's accuracy and token gains come from giving the retrieval layer an explicit model of entities and their relationships to reason over, instead of asking an LLM to reconstruct that structure from raw text at query time every single time. That is the same bet CCO and other mid-level ontologies make at a smaller, domain-general scale: disciplined, reusable entity and relation definitions pay for themselves once a system has to answer more than isolated lookup questions. A preview feature inside one vendor's retrieval stack does not settle how enterprises should architect that layer, but it is a visible data point that the default retrieval pattern for agentic systems is shifting toward explicit graph structure.