Unlock On-Prem Productivity with Agentic Retrieval in Foundry Local

In today’s connected world, customers expect instant, context-rich interactions, even in environments where cloud connectivity isn’t guaranteed. That’s where Retrieval-Augmented Generation at the edge comes in. Since we launched into public preview, we’ve watched teams across regulated, disconnected, and mission-critical environments push this technology into places cloud GenAI simply couldn’t reach. 

What we heard back shaped everything in this release: customers don’t just want retrieval. They want reasoning, they want agency, and they want an end-user experience that feels as natural as the one they already use in the cloud. 

Today at Build 2026, we're excited to introduce Agentic Retrieval, the next evolution of our on-prem RAG platform, enabled by Azure Arc and powered by Foundry language models. Agentic Retrieval is part of Microsoft's Adaptive Cloud approach, which extends Azure capabilities to wherever customer data and workloads actually live, with Edge AI focused on bringing reasoning and grounding to on-prem, distributed, and disconnected environments. Together with Foundry Local, Agentic Retrieval continues to shape Microsoft's Foundry Anywhere commitment: flexibility, resilience, and intelligence wherever customers operate. 

 

What’s new at Build 2026 

This release introduces three major pillars that work independently or together: 

• Agentic Retrieval engine: a first-party orchestration runtime for planning, reasoning, conversation state, and tool calls over your local data 

• Knowledge: a dedicated layer for organizing, curating, and governing your grounding data, exposed via MCP and connectable to any agentic retrieval layer 

• Chat UI: a production-ready, polished conversational experience that ships as the default UX for Agentic Retrieval and can also be deployed standalone 

Alongside, we’re delivering the platform upgrades customers asked for: flexible deployment modes (Agentic-only, Knowledge-only, or Combined), BYOM with pluggable backends, Foundry Local model catalog integration, Entra ID support, disconnected-ready, and hybrid search combined with agentic retrieval.

Agentic Retrieval: From Answering to Reasoning 

Classic RAG retrieves, then generates. Agentic Retrieval plans, reasons, and acts, running multi-step retrieval and tool invocation under a first-party orchestration runtime, entirely on your infrastructure. Under the hood it manages query planning, iterative multi-hop retrieval, tool calls via MCP, conversation state, and mandatory grounding with citations and audit logging built in. 

 

Image 1: Agentic reasoning trace: query decomposition, retrievals, tool calls, grounded answer

What customers can achieve: 

• Compliance, policy, and permit workflows for public sector, regulators, and defense operations, with data never leaving sovereign infrastructure 
• Multi-document synthesis across standards, technical manuals, contracts, and field procedures for industrial operators 
• An agentic chat experience for regulated and operational teams (engineers, inspectors, analysts) that reasons like a subject-matter expert 
• Auditable AI for sovereign and mission-critical environments, with every answer traceable to its source 

Knowledge: A First-Class, Governed Data Layer 

Great answers start with great knowledge. Knowledge is now a standalone component customers can deploy on its own or alongside Agentic Retrieval, exposed through an MCP wrapper so it can connect to any agentic retrieval layer, ours or yours. 

This release brings Collections (segmented groups of indexed knowledge with granular access permissions), multi-source ingestion across documents, tables, images, and SharePoint (indexed source moving to public preview), high-fidelity parsing for complex enterprise content, Bring Your Own MCP to connect customer-owned data sources directly into Agentic Retrieval and the chat experience, and governance enforced at the data layer itself. 

 

Image 2: Knowledge management view - collections, sources, and permission scopes

 

What customers can achieve: 

• Scope knowledge access to different slices of the same corpus, by plant, site, classification, or jurisdiction 
• Enforce data sovereignty, residency, and regulatory compliance at the knowledge layer itself 
• Ground both first-party Agentic Retrieval and BYO orchestration through a single governed source of truth across distributed sites 
• Keep classified, proprietary, and operational data fully on-prem while delivering premium chat experiences 
  

Chat UI: Production-Ready Conversational Experience 

Agentic Retrieval now ships with a polished, production-ready Chat UI as its default experience, and the same component can be deployed standalone for customers building their own stack on Foundry Local. 

Highlights include Entra ID authentication (MSAL login, Bearer tokens, user identity display), pluggable backends across AI Foundry, BYOM, or mock mode with zero code changes, Chain-of-Thought visibility and inline citations that make grounding transparent to end users, standalone frontend deployment via Helm chart and container image, and disconnected-ready operation for air-gapped environments. 

 

Image 3: Chat UI answer with Chain-of-Thought panel expanded and inline citations

What customers can achieve: 

• Deliver a polished end-user experience to operators, inspectors, and analysts without building UI from scratch 
• Build trust in regulated and industrial workflows through transparent, inspectable reasoning and grounding 
• Run the same UI across air-gapped facilities, sovereign clouds, and connected industrial sites 
• Accelerate rollout across public sector, defense, manufacturing, and other mission-critical environments 
  

Why This Release Matters 

Every update to our on-prem RAG platform has moved us toward a simple conviction: GenAI should be useful wherever customers operate, whether regulated or open, connected or disconnected, centralized or distributed. 

With Agentic Retrieval, Knowledge, and Chat UI coming together, backed by Foundry on Arc, BYOM, and fully disconnected support, this is no longer “cloud RAG, but local.” It’s an agentic knowledge platform purpose-built for the realities of enterprise data: on-prem, governed, and increasingly autonomous. 

Learn More 

• Explore Agentic retrieval documentation
• Read Foundry Local on Azure Local model inferencing blog post
• For more information reach out to the team at FoundryLocalOnAzure@microsoft.com