Foundry IQ adds MCP integration

Foundry IQ integrates with Foundry Agent Service via MCP to build enterprise AI agents that retrieve, reason over, and cite organizational data. It provides auto-chunking, vector embeddings, permission-aware retrieval, semantic reranking, and citation-backed responses for traceable, secure answers.

Foundry IQ now integrates with Foundry Agent Service via the Model Context Protocol. This enables AI agents to retrieve and reason over enterprise knowledge with permission controls.

Main feature/change and impact

Foundry IQ exposes an MCP endpoint that connects knowledge bases to agents. Agents call knowledge_base_retrieve to query unified indexes. The integration adds auto-chunking, auto-embedding, and semantic reranking to retrieval. Permission-aware retrieval enforces ACLs from SharePoint, OneLake, and Azure Blob Storage. The result is traceable, citation-backed responses aligned with organizational access controls.

Practical implications

Developers can create project connections using ProjectManagedIdentity to link agents to knowledge bases. Agents can decompose queries, run parallel searches, and rerank results with LLM assistance. Index maintenance is automated, reducing ingestion pipeline work. RBAC and Purview sensitivity labels are enforced at query time, ensuring responses respect data governance and audit requirements.

“You MUST use the knowledge_base_retrieve tool for every question. Include citations from sources.”

The integration requires creating a project connection, defining an MCP tool in the agent, and using the OpenAI client for conversations. Next steps are enabling RBAC on Azure AI Search and assigning the project managed identity appropriate roles. Implement the three-step setup and test agents with representative queries and datasets.

Key points from the article:

MCP connects agents to unified knowledge bases for grounded retrieval.

Auto-chunking and auto-embedding simplify ingestion pipelines.

Permission-aware retrieval enforces source ACLs at query time.

Agentic retrieval decomposes queries and performs parallel searches.

Semantic reranking and citations improve relevance and traceability.

Related Coverage:

• We’re bringing our growing MAI model family to every developer in Foundry, including: MAI-Transcribe-1, most accurate transcription model in world across 25 languages; MAI-Voice-1, natural, expressive speech generation; MAI-Image-2, our most capable image
• oBeaver — A Beaver That Runs LLMs on Your Machine 🦫
• Learn how to build agents and workflows in Python

From the Microsoft Developer Community Blog articles

---