Edit

What's new in Agentic Retrieval in Foundry Local

This article lists recent features and improvements in Agentic Retrieval (formerly Edge RAG enabled by Azure Arc).

Important

Agentic Retrieval in Foundry Local is currently in PREVIEW. See the Supplemental Terms of Use for Microsoft Azure Previews for legal terms that apply to Azure features that are in beta, preview, or otherwise not yet released into general availability.

July 2026

This month introduces improvements for ingestion, agentic conversation quality, security posture, and deployment resilience.

Release of extension version 0.9.5

This release focuses on ingestion scale and reliability, agentic memory management for long conversations, security hardening, and a smoother, more secure deployment experience.

Faster, more reliable ingestion
Document parsing now scales horizontally across multiple workers with automatic retries, which reduces failures for large and complex documents. The process reports unsupported file formats as skipped instead of silently dropping them. Parsing reports coverage to help you detect missing content. You can cancel ingestion runs cleanly, and the process handles duplicate or stuck-job edge cases without failing.

Agentic memory for long conversations
Automatic context and memory compaction keeps long, tool-heavy agentic conversations within the model's context window, using a strategy tuned for small local models. The full conversation history is preserved while the active context is compacted automatically, so multistep conversations stay reliable.

Security hardening
Ongoing security hardening includes dependency and vulnerability patching, tighter handling of secrets and tokens so they don't leak into logs or telemetry, and cleaner credential handling during deployment.

More secure networking and deployment
For secure network segmentation, the cluster needs a CNI that enforces Kubernetes NetworkPolicy (Calico recommended), and the installer checks for this requirement. Agentic Retrieval ships and manages its own ingress controller, so bring-your-own ingress isn't supported. Deployment is smoother and more resilient, with GPU/A100 support and automatic prerequisite setup and cleanup.

Recommended 32K context window
Configure the Foundry Local model with a 32K token window to improve retrieval and tool-use quality.

Agentic retrieval and evaluation improvements
Connect multiple knowledge sources to a single knowledge base, use more flexible Chat API inputs with clearer inference error reporting, and measure answer quality across conversation turns with multistep evaluation.

Improved disconnected operations
This release improves support for disconnected deployments, including SharePoint integration and simplified extension packaging. For more information, see Download and import the Agentic Retrieval expansion pack.

June 2026

Agentic Retrieval (preview)

This release renames Edge RAG enabled by Azure Arc to Agentic Retrieval, a retrieval-augmented generation platform at the edge, and adds an agentic layer for AI agent orchestration. You can build agents that manage multistep conversations, invoke external tools, and ground responses in your on-premises data.

Agentic layer

Build multithreaded AI agents that orchestrate across knowledge sources and external tools:

  • Agents Runtime — Create threads, send messages, execute runs with streaming (SSE). OpenAI Assistants-compatible API.
  • Knowledge Base Manager — Manages the default knowledge base (GET, PATCH, PUT). Knowledge bases group knowledge sources and define what data the system can access.
  • Knowledge Sources — Register MCP server connections as self-contained knowledge sources (two kinds: remote_mcp and indexed_sources_mcp), and organize them into knowledge bases.
  • Built-in MCP Server — Six search tools (hybrid, vector, text, image, multimodal, list collections) exposed over Model Context Protocol for any MCP-compatible client.
  • Collections — Multiple named collections for vector data with per-collection RBAC, replacing the single-index model.

Independent Deployment Modes

Deploy exactly what you need with the new layerSelection parameter:

  • combined (default) — full platform
  • agentic — agents without local data ingestion (no GPUs needed)
  • knowledge — data ingestion and RAG without agent orchestration

Foundry Local or BYOM endpoints required

Phi-3.5 and Mistral-7B are no longer bundled with the extension. All deployments now require a language model endpoint. Use a Foundry Local endpoint on Azure Local as the recommended option, or use an external BYOM endpoint that supports an OpenAI-compatible chat completions API, such as one deployed in Microsoft Foundry.

GPU Count Reduced

GPU requirement reduced from 4 to 2 (embedding, image). Docling now runs on CPU. The LLM runs externally via BYOM.

Seven REST API references published

Complete API documentation for: Agents Runtime, Knowledge Base Manager, Knowledge Sources, Collections, MCP Server, Ingestion, and Inference.

For more information, see the following articles:

February 2026

Release of extension version 0.8.5

This release focuses on reliability, stability, and security improvements for Edge RAG. These updates make the system more reliable and predictable, with better error handling and enhanced security protections.

Enhanced search reliability
Search operations are more dependable with improved deep search accuracy for complex queries. You get clearer error messages that tell you whether the system is unavailable or the knowledge base is empty, so you know exactly what's happening and can take the right action.

More reliable data ingestion
Data ingestion now handles edge cases without failing, including files without extensions and empty folders. The system reports which files it skipped and why, giving you full visibility into the ingestion process. NFS connections validate accessibility, sockets, and permissions upfront, with clear error messages when problems occur.

Improved system reliability
Configurable timeouts for document conversion, SharePoint connections, LLM inference calls, and database operations prevent indefinite hangs and resource exhaustion. Operations fail predictably with clear diagnostics, so you can tune settings for your environment.

Better troubleshooting
Deployment failures now provide accessible logs, so you can identify root causes faster without escalating to support. This improvement reduces time to resolution and helps you get Agentic Retrieval running smoothly.

Critical security updates
This release addresses two important vulnerabilities: a critical Next.js Denial-of-Service vulnerability (GHSA-5j59-xgg2-r9c4) and a high-severity Langchain XXE vulnerability that could allow unauthorized file access. Dependency updates also resolve version conflicts and improve overall system stability.

November 2025

Release of extension version 0.8.2

This release of Edge RAG introduces several new features, enhancements, and improvements designed to expand capabilities, improve performance, and streamline the user experience.

Deep search
Find the most relevant information with the new deep search model. Deep search uses production-ready LazyGraph RAG with high-quality RAG inferencing. Edge RAG now explores and connects data across sources at query time, so you get comprehensive answers without heavy upfront processing. For more information, see Search types in Agentic Retrieval.

High-fidelity parsing
Choose between basic text extraction or advanced parsing to capture tables, images, and more. By using advanced parsing, Edge RAG offers OCR-enabled support for documents, tables, and images. Tailor data ingestion to your needs for more accurate results. For more information, see Advanced data parsing for Agentic Retrieval.

Performance and scale
Experience up to 5× faster query performance for hybrid search and 100× faster ingestion of live-streamed images from the previous Edge RAG extension version 0.1.5.

Advanced search and chat experience

Edge RAG now offers a more capable and flexible search and chat experience, making it easier to find information and interact with your data through new capabilities and interface improvements.

  • Use hybrid multimodal search to retrieve images and deliver responses with relevant images. For more information, see Search types in Agentic Retrieval.
  • Get markdown-formatted responses that support images and formatted text for responses that are easier to read and interpret.
  • Chat directly with the language model, without using your organization’s data as context. Use the model only option to ask general questions, test the model’s capabilities, or get responses that aren’t influenced by your ingested data. Switch between knowledge-based chat and model-only chat to fit your needs. For more information, see Knowledge layer configuration.

Preview support for disconnected scenarios

Edge RAG is supported as part of a preview for disconnected operations for Azure Local.

Related content

For more information about this release, see:

Related content

  • Last updated on