Foundry IQ prerequisites
This is the shared baseline for every Foundry IQ knowledge source in GPT-RAG. Read this once before you enable your first Foundry IQ knowledge source. Each per-source how-to page lists only its own extra prerequisites on top of what is covered here.
If you have not read the Grounding sources overview yet, start there. It explains what a Knowledge Base and a Knowledge Source are and how GPT-RAG uses them.
Preview API
Foundry IQ uses the Azure AI Search agentic retrieval API. GPT-RAG pins
the 2026-05-01-preview version. Preview features are provided without a
service-level agreement and may change without notice. Do not depend on
them for production workloads until they are generally available. See the
Microsoft preview terms.
What GPT-RAG does for you
azd provision on GPT-RAG v3.0.2+ with AI Landing Zone v2.1.2+ handles the
common Foundry IQ setup: it deploys a Foundry project, an Azure AI Search
service with the knowledgeRetrieval billing plan enabled, assigns the RBAC
roles listed below, creates the Knowledge Base, and seeds all the Foundry IQ
App Configuration keys under the gpt-rag label.
If you used the default deployment path, most of this page is background information. The parts you may still need to act on yourself are:
- Region selection before
azd provision. - Feature flags gated by Microsoft (Work IQ, Web).
- Per-source authorization steps for on-behalf-of (OBO) knowledge sources.
- Choosing whether to enable a query planning LLM on the Knowledge Base.
Everything else on this page is what your deployment already has.
Before you start
- An Azure subscription where you will deploy GPT-RAG.
- Local roles on that subscription:
ContributorandUser Access Administratorfor the identity that runsazd provision. This is the deploy guide baseline.
- A region that supports Azure AI Search agentic retrieval. See Azure AI Search regions list and pick a region with the "Agentic retrieval" column set. GPT-RAG's regional preflight will flag your choice before deployment starts.
- Azure AI Search tier: Basic or higher. The Free tier does not support managed identity, which Foundry IQ needs to read the Knowledge Base and optionally to call the query planning LLM. See tier and pricing.
Foundry project and preview API
- A Foundry project in the same subscription and region as your Search service. GPT-RAG deploys one for you through AI Landing Zone. To create one manually, see Create a Foundry project.
- Preview API version
FOUNDRY_IQ_API_VERSION=2026-05-01-preview. This is required for per-user permissions on native knowledge sources and forfilterAddOnon thesearchIndexknowledge source. See Foundry IQ documents for the full key list.
RBAC and permissions
Two distinct identities need permissions for Foundry IQ to work end to end:
- The executor identity that provisions and manages Knowledge Bases and Knowledge Sources.
- The runtime identities that call the Knowledge Base at query time: the GPT-RAG orchestrator managed identity and the Foundry project managed identity.
GPT-RAG's provisioning assigns these roles automatically. The table is here so you can verify them in the Azure portal or reproduce them in a locked-down subscription where you deploy manually.
| Assignee | Role | Scope | Why |
|---|---|---|---|
Executor (user or SP running azd provision) |
Search Service Contributor |
Azure AI Search service | Create and update the Knowledge Base and Knowledge Sources. |
| Executor | Search Index Data Contributor |
Azure AI Search service | Populate indexes generated by indexed knowledge sources. |
| Executor | App Configuration Data Owner |
App Configuration store | Write the gpt-rag-labeled keys. |
| Foundry project managed identity | Search Index Data Reader |
Azure AI Search service | Read Knowledge Base results. |
| Foundry project managed identity | Search Service Contributor |
Azure AI Search service | Create the AI Search connection used by the Knowledge Base. |
| Foundry project managed identity | Storage Blob Data Reader |
GenAI storage account | Read the documents container for the default Blob path. |
| Azure AI Search service managed identity | Cognitive Services User |
Foundry account | Call the Foundry model account (vectorizers and, if configured, the query planning LLM). Assigned by default. |
| Orchestrator container app managed identity | Search Index Data Reader |
Azure AI Search service | Query the Knowledge Base at runtime. |
| Orchestrator container app managed identity | Cognitive Services OpenAI User |
Foundry account | Call the model that generates the final answer. |
Full role tables for every GPT-RAG component are in Deployment guide: role assignments.
For the underlying Microsoft-documented roles, see Create a Knowledge Base: prerequisites and Connect a Foundry IQ Knowledge Base to Foundry Agent Service: authentication and permissions.
Authorization model per knowledge source
Foundry IQ has two authorization patterns. Which one applies determines whether the source can serve anonymous chat and whether your users need to sign in.
| Knowledge source kind | Auth pattern | Signed-in user required |
|---|---|---|
azureBlob (Documents, default) |
App-only. Foundry IQ reads Blob using its own identity, with optional rbacScope or Purview labels for trimming. |
No. |
searchIndex (custom ingestion path) |
App-only. Optional filterAddOn for GPT-RAG security fields. |
No. |
indexedOneLake |
App-only. Foundry IQ managed identity reads the lakehouse. | No. |
indexedSharePoint |
App-only. Microsoft Graph Sites.Selected or Sites.Read.All on a registered Entra app. No per-user ACL trimming in this preview. |
No. |
web (Grounding with Bing) |
App-only. No ACL. Billed per call. | No. |
mcpServer (generic MCP server) |
Registration is app-only. Endpoint authentication is separate and requires a compatible runtime for query-time headers. App-only managed identity and per-user OBO are distinct runtime token-acquisition options. See Generic MCP server. | Only if you choose OBO for per-user endpoint authorization. |
workIq |
On-behalf-of (OBO). Orchestrator forwards the user's delegated token in x-ms-query-source-authorization. |
Yes. |
remoteSharePoint |
OBO through the Microsoft 365 Copilot Retrieval API. | Yes. |
fabricOntology |
OBO to Microsoft Fabric. | Yes. |
fabricDataAgent |
OBO to Microsoft Fabric. | Yes. |
For every OBO knowledge source, the orchestrator front end must be configured to forward a delegated token. This is the same plumbing that Auth and Doc Security describes for GPT-RAG's own security-trimmed retrieval. Each per-source page lists the specific tenant consent, delegated permissions, and licensing you need.
For mcpServer, do not confuse the orchestrator's managed identity with an
OBO token. Managed identity is app-only and gives every user the same MCP
authorization scope. OBO preserves the signed-in user's identity and only
works when the MCP endpoint validates that delegated token. The optional
queryHeaders metadata selects how a compatible orchestrator resolves the
header and requires an explicit, non-secret scope for either mode. The
provisioning change stores this metadata in App Configuration but never
renders it into Search registration or retrieve parameters. Runtime
resolution still requires a compatible orchestrator. See
Generic MCP server.
Feature flags gated by Microsoft
Two knowledge sources require Microsoft-controlled enrollment before you can use them. GPT-RAG cannot enable these for you.
| Knowledge source | Gate | Where to enable |
|---|---|---|
Work IQ (workIq) |
Gated public preview. Requires the EnableFoundryIQWithWorkIQ subscription feature and one-time tenant admin consent for the Work IQ service principal. |
See Work IQ prerequisites. |
Web (web) |
Requires the Microsoft.Search/WebKnowledgeSource subscription feature. Billed per Bing call. |
See Web grounding prerequisites. |
Optional: query planning LLM
The Knowledge Base can call an Azure OpenAI model to plan queries and choose
which knowledge sources to hit. GPT-RAG does not require you to configure
this: at retrievalReasoningEffort=minimal (the current default) the
Knowledge Base skips LLM planning entirely. Configure a planning model when:
- You enable the
webknowledge source. Web requires an LLM. - You enable one or more generic MCP server (
mcpServer) sources. MCP requireslowormediumreasoning so the planner can select a tool and generate its arguments. A compatible orchestrator must also use amessages-based retrieve request; see Generic MCP server. - You want
lowormediumreasoning effort so the Knowledge Base can select and blend across multiple sources.
Supported planning models on 2026-05-01-preview include gpt-4o,
gpt-4o-mini, gpt-4.1, gpt-4.1-mini, gpt-4.1-nano, gpt-5,
gpt-5-mini, gpt-5-nano, gpt-5.1, gpt-5.2, and the gpt-5.4 family.
The authoritative list moves; check
Supported models
before pinning one.
The GPT-RAG orchestrator uses its own model for the final answer, configured
through CHAT_DEPLOYMENT_NAME in App Configuration. That is separate from the
Knowledge Base planning model. See
Azure Direct Models for how to change it.
Configuration
Every Foundry IQ toggle lives in Azure App Configuration under the
gpt-rag label, or as container environment variables when the deployment
runs in containerEnv mode. GPT-RAG seeds the shared keys and each source's
per-source keys, so you rarely start from an empty store. The lever that
picks the default Blob path versus the custom searchIndex ingestion path
is documented in
Foundry IQ: Documents.
Common shared keys (defaults set by GPT-RAG):
| Key | Default | Purpose |
|---|---|---|
RETRIEVAL_BACKEND |
foundry_iq |
Select the Foundry IQ backend. Set to ai_search to roll back. |
KNOWLEDGE_BASE_NAME |
Generated | Foundry IQ Knowledge Base name. |
KNOWLEDGE_BASE_ENDPOINT |
Generated | Azure AI Search endpoint that hosts the Knowledge Base. |
KNOWLEDGE_BASE_CONNECTION_ID |
Generated | Foundry connection ID dedicated to the Knowledge Base. |
FOUNDRY_IQ_API_VERSION |
2026-05-01-preview |
Preview API version required for per-user permissions and filterAddOn. |
FOUNDRY_IQ_KNOWLEDGE_RETRIEVAL_BILLING_PLAN |
free |
free or standard for the Azure AI Search agentic retrieval billing plan. |
FOUNDRY_IQ_MAX_OUTPUT_DOCUMENTS |
5 |
Cap on documents returned by the Knowledge Base per call. |
FOUNDRY_IQ_MAX_RUNTIME_SECONDS |
120 |
Per-request retrieval budget. OBO sources such as Work IQ can approach this budget. |
Each per-source page documents its own keys in full.
Verify your baseline
Before you enable any Foundry IQ knowledge source, confirm the four items below on your deployment.
RETRIEVAL_BACKEND=foundry_iqis set in App Configuration under thegpt-raglabel.FOUNDRY_IQ_API_VERSION=2026-05-01-previewis set.- The Foundry project, the Azure AI Search service, and the GenAI storage account are all in the same subscription and Entra tenant. In the Azure portal, open each resource and check the Subscription and Directory (tenant) properties match.
- The Search region shows "Agentic retrieval" in the Azure AI Search regions list.
If any of those are missing, fix them before you turn on a knowledge source.
Related reading
- Grounding sources overview
- Foundry IQ: Documents
- Deployment guide
- Auth and Doc Security
- Microsoft Learn: What is Foundry IQ?
- Microsoft Learn: What is a knowledge source?
- Microsoft Learn: Create a Knowledge Base