Skip to content

What's New

This page highlights the notable features added to GPT-RAG over time. It is a feature outline, not a full changelog. For the detailed record of every release, patch, and fix, see the GitHub releases.

📌 Check out what's coming next (Azure org only)

July 2026

  • Governance audit and provenance integration (GPT-RAG v3.7.0). The umbrella release pins orchestrator v3.8.0 and ingestion v2.5.0 to the shared audit-event-v1 contract. Audit, sensitive-content capture, actor pseudonymization, and ingestion provenance remain off by default. When no operator-managed HMAC reference exists, post-provisioning creates a stable 256-bit key in Key Vault and registers only its Key Vault reference; existing Key Vault references are preserved. It also adds optional provenance fields to existing Azure AI Search indexes without recreating them. See Audit Contract v1 for configuration, migration, rollback, and evidence limitations. delete_after records policy intent only; it does not trigger automatic deletion.
  • Generic MCP Server knowledge source (preview, v3.6.0). The Knowledge Base can call approved tools on a remote MCP server you operate or trust and blend those results with other Foundry IQ sources. The feature is off by default, requires a trusted-host allowlist, and must be deployed together with orchestrator v3.7.0. See Foundry IQ: Generic MCP server.
  • Four new Foundry IQ knowledge sources (preview, v3.5.0). All off by default and gated by their own App Config flags.
    • SharePoint remote (remoteSharePoint). Live retrieval from SharePoint via the Copilot Retrieval API, with per-user OBO ACL. See Foundry IQ: SharePoint remote.
    • OneLake indexed (indexedOneLake). Fabric OneLake as a knowledge source, indexed and queried natively by Foundry IQ. See Foundry IQ: OneLake.
    • SharePoint indexed (indexedSharePoint). Foundry IQ indexes a SharePoint site using app-only Microsoft Graph auth (Sites.Selected preferred). See Foundry IQ: SharePoint indexed.
    • Web grounding (web). Public web results scoped by allow / block domain lists, billed per Bing call. See Foundry IQ: Web grounding.
  • Fabric Data Agent grounding (preview, v3.4.3). Answers can now be grounded on a Microsoft Fabric Data Agent, a virtual analyst that runs queries over Fabric data. It sits alongside documents, Work IQ, and the Fabric ontology source. Off by default, same OBO auth and data-egress notes as the other Fabric source. See Foundry IQ: Fabric Data Agent.
  • App Configuration seeding fix (v3.4.1). azd provision now seeds the WORK_IQ_* and FABRIC_IQ_* App Configuration keys with enabled=false and empty string defaults, so operators no longer have to create them by hand before flipping Work IQ or the Fabric knowledge sources on in an existing environment.
  • Fabric ontology grounding (preview, v3.4.0). Answers can now blend analytical data from a Microsoft Fabric ontology alongside documents and Work IQ. Off by default, gated by data residency (may route data outside your Foundry region). See Foundry IQ: Fabric ontology and the Grounding sources overview.
  • Work IQ grounding (public preview). The orchestrator can now augment retrieval with signals from the signed-in user's Microsoft 365 world (mail, meetings, files, chats, people) through Foundry IQ. Off by default. Requires a signed-in user, an M365 Copilot license, and gated preview access. Work IQ, the Fabric knowledge sources, and the documents source can all run together on the same Knowledge Base. See Foundry IQ: Work IQ.
  • CAF resource naming by default. Fresh Basic and Zero Trust deployments produce Cloud Adoption Framework-aligned resource names automatically, with an opt-out to the legacy scheme. See resource naming guide.
  • Admin dashboard sign-in. The orchestrator dashboard SPA now performs its own Microsoft Entra ID sign-in using MSAL (Authorization Code + PKCE), with a clear "signed in but missing Admin role" state and a runtime auth-config endpoint so the SPA can bootstrap in either authenticated or open mode. See Admin Dashboard Sign-in.

June 2026

  • Foundry IQ retrieval backend. New deployments retrieve through a native Azure Blob Knowledge Source in Foundry IQ by default, with Azure AI Search still fully supported for existing deployments and custom ingestion pipelines. See Grounding sources overview.
  • Operator dashboards. An opt-in dashboard at /dashboard on the orchestrator and ingestion apps gives admins conversation analytics, a conversations browser, live configuration editing, and on-demand job runs with queue and schedule visibility.

May 2026

  • In-chat file uploads. Users can upload files directly in a conversation, and the uploader identity is preserved end to end so uploaded content is retrieved only within that conversation.
  • Existing-platform deployment. Operators can integrate with shared Private DNS zones, hub networking, observability, jumpbox/Bastion/NAT, and other existing platform resources from the root parameters file, without editing the infra submodule.
  • AI Landing Zone v2. Adds uniform IP allow-listing across services, bring-your-own DNS/observability/hub-and-spoke parameters, and pre-flight validation that fails fast on region, quota, or parameter problems before reaching ARM.

April 2026

  • Ingestion Admin Dashboard with Content Understanding. A dashboard for monitoring and managing ingestion jobs and files, plus document analysis via Azure AI Foundry Content Understanding by default for roughly 69% lower cost per page.
  • Conversation history. List, resume, and delete past conversations from a sidebar in the chat UI, with inline multimodal images between response steps.

March 2026

  • Orchestration strategies. Choose between Azure AI Foundry Agent Service v2, the Microsoft Agent Framework, a combined mode, and multimodal, with Bicep infrastructure moved to the external AI Landing Zone module.

January 2026

  • Document-level security. Microsoft Entra ID authentication with orchestrator-side identity validation and RBAC/ACL enforcement in retrieval, so access control is applied end to end. See Authentication and Document-Level Security.

December 2025

  • Azure Direct Models (Microsoft Foundry). Use Foundry "Direct from Azure" models such as Mistral, DeepSeek, Grok, and Llama through the Foundry inference APIs with Entra ID authentication. See Azure Direct Models.
  • SharePoint Lists. The SharePoint connector covers both document libraries and generic lists, keeping your index in sync with list items and documents. See SharePoint Data Source.

October 2025

  • Bring Your Own VNet. Deploy GPT-RAG inside your existing virtual network with full control over network boundaries, DNS, and routing.
  • Agentic retrieval. Agent-driven retrieval that dynamically selects and combines information sources for more grounded responses.

September 2025

  • User feedback loop. End users can rate assistant responses thumbs-up or thumbs-down, stored alongside conversation history to improve quality. See User Feedback Configuration.
© 2025 GPT-RAG — powered by ❤️ and coffee ☕