Note
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
Agentic Retrieval helps you deploy Retrieval Augmented Generation (RAG) solutions at the edge. This article lists new features, improvements, and important changes for each release. Use these notes to plan, deploy, and manage Agentic Retrieval in your organization.
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
Extension version: 0.9.5 [Preview]
This release focuses on ingestion scale and reliability, agentic memory management for long conversations, security hardening, and a smoother, more secure deployment experience. For current limitations and workarounds, see Known issues in Agentic Retrieval in Foundry Local.
Ingestion
| Feature | Description |
|---|---|
| Faster, more reliable large-scale ingestion | Document parsing scales horizontally across multiple workers, with automatic retries, so large and complex documents are far less likely to fail. |
| Clearer visibility into parsing | Files in unsupported formats are reported as skipped instead of being silently dropped, and parsing reports coverage so any missing or dropped content is detectable. |
| More robust job handling | Ingestion runs can be cancelled cleanly, and duplicate or stuck-job edge cases are handled gracefully. |
Agentic memory and long conversations
| Feature | Description |
|---|---|
| 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
| Feature | Description |
|---|---|
| Ongoing security hardening | 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. |
Networking and deployment
| Feature | Description |
|---|---|
| NetworkPolicy-capable CNI required | For secure network segmentation, the cluster needs a CNI that enforces Kubernetes NetworkPolicy (Calico recommended). The installer checks for this up front and fails early with guidance if it's missing. |
| Bring-your-own ingress controller isn't supported | Agentic Retrieval ships and manages its own ingress-nginx controller. For more information, see Known issues in Agentic Retrieval in Foundry Local. |
| Smoother, more resilient deployment | Includes GPU/A100 support and automatic prerequisite setup and cleanup. |
Foundry Local model configuration
| Feature | Description |
|---|---|
| Recommended context window: 32K | Configure the Foundry Local model with a 32K token window for the best results with retrieval and tool use. |
Agentic retrieval and evaluation
| Feature | Description |
|---|---|
| Multiple knowledge sources per knowledge base | Connect multiple knowledge sources to a single knowledge base. |
| More flexible Chat API inputs | Accepts more flexible inputs, with clearer and more accurate inference error reporting. |
| Multi-turn evaluation | Measure answer quality across conversation turns. |
Disconnected operations on Azure Local
| Feature | Description |
|---|---|
| Improved disconnected support | Improved support for disconnected and air-gapped deployments, including SharePoint integration and simplified extension packaging. For more information, see Download and import the Agentic Retrieval expansion pack. |
June 2026
Extension version: 0.9.3 [Preview]
Changes
| Change | Details |
|---|---|
| Product renamed | Edge RAG enabled by Azure Arc is now Agentic Retrieval. The Azure extension type remains microsoft.arc.rag. |
| SLM models removed | Phi-3.5 and Mistral-7B models are no longer bundled. Bring Your Own Model (BYOM) is the only model path. You must provide an external LLM endpoint. |
| GPU count reduced | 4 → 2 GPUs required. GPUs are used for text embedding (BGE-M3) and image embedding (CLIP ViT-L/14). Docling (document parser) now runs on CPU. The LLM runs externally via BYOM. |
| BYOM is now mandatory | The model endpoint step is no longer optional. All deployments require an OpenAI-compatible LLM endpoint. |
New features
| Feature | Description |
|---|---|
| Independent layer deployment | New layerSelection ARM parameter: combined (default), agentic, knowledge. Deploy layers independently. |
| Agents Runtime API | Full agent orchestration with threads, messages, runs, and streaming (SSE). OpenAI Assistants-compatible API. Port 8080. |
| Knowledge Base Manager API | Manages the default knowledge base (GET, PATCH, PUT). Port 8080. |
| Knowledge Sources API | Register MCP server connections as self-contained knowledge sources (two kinds: remote_mcp and indexed_sources_mcp). Port 3005. |
| Collections API | Manage multiple collections for vector data with per-collection RBAC. Port 3002. |
| MCP server | Built-in MCP server with 6 search tools over Model Context Protocol. Port 8080. |
| Ingestion API | Public REST API for programmatic document ingestion. Port 8000. |
| Inference API | Public REST API for RAG chat completions and model-only queries. Port 3001. |
| Agentic chat UI | New agentic chat interface with multi-turn conversations, thread history, and streaming. |
February 2026
Extension version: 0.8.5 [Preview]
This release includes bug fixes, reliability improvements, and critical security updates.
Search and query
This release improves the accuracy and reliability of search operations.
- Deep search returns accurate results for complex queries, including time-based and count-based questions.
- Timeouts for LLM inference calls, Vector DB queries, and database operations prevent resource exhaustion.
- Error messages clearly show whether the system is unavailable or the knowledge base is empty.
Data ingestion
Data ingestion is more robust and handles edge cases without failure.
- The system skips files without extensions and reports which files it skipped and why.
- You can now ingest to empty folders or folders with mixed content without failures.
- NFS connection validation checks accessibility, socket connections, and permissions, and provides clear error messages when problems occur.
- You can configure timeouts for document conversion, SharePoint connections, and file processing to prevent indefinite hangs.
Diagnostics and observability
Improved log access makes troubleshooting deployment issues faster. You can access logs during and after deployment failures, so you can identify root causes faster.
Security updates
This release addresses critical and high-severity vulnerabilities.
- This release fixes a critical Next.js denial-of-service vulnerability (GHSA-5j59-xgg2-r9c4) through an upgrade to version 14.2.35.
- This release resolves a high-severity Langchain XXE vulnerability that could allow unauthorized file access during document parsing.
- Dependency updates resolve version conflicts and improve system stability.
November 2025
Extension version: 0.8.2 [Preview]
The following table summarizes the updates included in this release.
| Capability | Description |
|---|---|
| Data ingestion | - Improved parsing for documents, including tables and charts. - Added new ingestion type options for high-fidelity document processing. For more information, see Advanced data parsing for Agentic Retrieval |
| Data query and chat | - Added deep search model search type for production-class LazyGraph RAG with industry-leading RAG inferencing quality. - Added hybrid multimodal search with image retrieval and image-rich outputs. - Added support for responses in markdown for improved readability. - Added a model-only chat option for direct model chat. For more information, see Knowledge layer configuration and Advanced data parsing for Agentic Retrieval. |
| Performance enhancements | From previous release version: - Achieved a 5× faster query performance for hybrid search. - Delivered 100× faster ingestion of live-streamed images. |
| Preview support for disconnected scenarios | Agentic Retrieval supported as part of a preview for disconnected operations for Azure Local. |
For more information, see What's new in Agentic Retrieval.
July 2025
Extension version: 0.1.5 [Preview]
- Resolved authentication problems for endpoints that Microsoft Foundry created for bring your own model (BYOM) deployments.
- Temporarily disabled chat history to improve performance. Each question is answered based only on retrieved content.
- Improved security.
May 2025
Extension version: 0.1.3 [Preview]
Edge RAG enabled by Azure Arc is available as a public preview.
See the related blog posts:
- Transforming On-Premises Data with RAG Capabilities on Azure Local
- Empowering the Physical World with AI - Unlocking AI at the Edge with Azure Arc
- Unlocking AI Apps Across Boundaries with Azure