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 has multiple network boundaries. Protecting the agent-to-Search connection doesn't automatically protect Search connections to data, models, or remote knowledge sources. For Foundry IQ (Azure AI Search) architectures, evaluate each traffic path separately before restricting public access.
Choose a supported topology based on every required traffic path. Use public endpoints with access controls, or combine private inbound Search access with explicitly permitted public dependencies. For eligible dependencies, configure supported private paths.
This article explains those boundaries and links to the existing setup and support requirements. It doesn't imply that every provider, knowledge source, or Foundry IQ service supports an entirely private deployment.
Network boundaries and defaults
Separate the control plane from the data plane:
- Control plane: Azure Resource Manager manages Foundry IQ services, identities, private endpoints, and shared private link resources through
management.azure.com. Management permissions and successful resource creation don't prove that a runtime data path is private. - Data plane: Clients call the Search endpoint, typically
<service-name>.search.windows.net, to manage search objects and retrieve content. Search can then call source, model, and other dependency endpoints.
Treat a public-endpoint connection as public unless a specific supported private path is configured and, where required, approved. An endpoint firewall, trusted-service exception, or network security perimeter access rule isn't equivalent to a Private Link connection.
Managed identity supplies an identity, and RBAC authorizes operations. Neither establishes network isolation. Conversely, a private endpoint doesn't replace authentication or data permissions.
Inbound access to Search
The client or agent runtime initiates the connection to Search. An inbound private endpoint gives Search a private IP address reachable from the connected network. DNS and network reachability must work from the actual caller, not just from an administrator's computer.
For Foundry-hosted agents, follow the private inbound connectivity tutorial for the applicable deployment. A private endpoint in your own virtual network doesn't by itself connect every hosted agent runtime.
Keep Search authentication and roles separate from that network setup. For Microsoft Entra authentication, Search uses the https://search.azure.com/.default token scope. Source-specific user tokens used for permission enforcement don't replace authentication to Search.
Inbound access and outbound access are independent. S3 HD and Serverless outbound shared private link restrictions don't imply that inbound private endpoints to Search are unsupported. See tier and workload limits.
Indexing-time outbound dependencies
Indexed knowledge sources query a local Search index after ingestion. Their ingestion path can still depend on external data and model services. Bring-your-own indexes use your ingestion pipeline; generated indexed knowledge sources use service-created resources. File knowledge sources use a separate direct-upload path.
An outbound Search shared private link isn't the inbound Search private endpoint. On eligible configurations, Search creates the connection to a supported target subresource, the target owner approves it, and indexer execution must use the supported private environment.
The shared private link prerequisites and supported resource types own tier, region, service-age, target, and group-ID requirements. Use those requirements rather than assuming that support for one Storage or Foundry endpoint covers another.
S3 HD (standard3 with hostingMode: HighDensity) and Serverless indexers run only in the multitenant environment and don't support outbound Search shared private links to sources, skills, or models. Don't set executionEnvironment: private to bypass this restriction. For unsupported legacy S3 HD resources, use the workload-aware remediation procedure.
Traffic paths and setup owners
The following map identifies dependencies to inventory. It isn't a second resource support matrix. Each linked owner defines its supported endpoints, identities, permissions, versions, and network configuration.
| Traffic path | Initiator and destination | Requirements and setup owner |
|---|---|---|
| Source ingestion | Search indexer to the configured data source endpoint. | Use the connector requirements, managed identity guidance, and indexer network controls. Indexed knowledge-source private configuration has its own source-specific requirements. |
| Embedding and model skills | Search processing to the configured model or custom skill endpoint. | Check the skill's identity, target resource, and endpoint requirements. Use supported shared private link targets only on eligible tiers and workloads. |
| Built-in skill billing connection | Search to the attached Foundry resource for keyless billing. | This is distinct from a skill's processing path. Follow keyless skill billing and its network prerequisites. |
| Auxiliary storage | Search processing to storage for enrichment caches, debug state, or knowledge stores. | Inventory these separately from source storage. See network-protected storage and feature-specific limitations, including debug sessions and private connectivity. |
| Image assets | Search ingestion and retrieval processing to the configured asset store. | Follow image-serving requirements. Don't assume that protecting the source container also protects asset storage or every model call that uses an image. |
Query-time models and remote knowledge sources
Query-time dependencies differ from ingestion. A knowledge base can invoke a configured model for query planning or answer synthesis. A vector query can also invoke a configured vectorizer. Protecting source ingestion doesn't protect these calls.
| Traffic path | Initiator and destination | Requirements and setup owner |
|---|---|---|
| Query planning and answer synthesis | Search knowledge base to its configured model deployment. | Follow knowledge base model access requirements. For supported private model connections, check shared private link eligibility and targets. |
| Query vectorization | Search vectorizer to its configured embedding endpoint. | Follow the vectorizer's provider requirements and supported private targets. Indexing-time embedding configuration doesn't establish query-time routing. |
| Remote knowledge retrieval | Search retrieval engine to an external platform or tool endpoint. | Remote content isn't ingested into a local Search index. Use each provider's authentication and endpoint requirements; don't assume that indexer shared private links apply. |
Remote sources have different access models:
- Remote SharePoint retrieves content on behalf of the user. SharePoint Online uses public endpoints and isn't a Search shared private link target.
- Work IQ queries live Microsoft 365 data on behalf of the user. Search exchanges the supplied user token for a delegated Work IQ token; follow the provider's authentication and endpoint requirements.
- Fabric data agents and Fabric ontology have their own provider and authorization requirements.
- Web knowledge sources access web content. Don't treat this as a private source-ingestion path.
- MCP server knowledge sources call configured external tool endpoints with their supported authentication.
For each selected provider, confirm the supported network route before disabling its public endpoint. A tenant-scoped identity or successful authenticated call doesn't prove a private route. Don't infer uniform fallback or fail-closed behavior across these providers.
Choose a supported topology
Choose a topology based on every required dependency, not just the Search endpoint:
- Public endpoints with access controls: Use supported authentication, authorization, firewalls, and applicable perimeter rules. These controls don't turn public endpoints into private endpoints.
- Mixed topology: Protect inbound Search access privately while retaining explicitly permitted public paths for dependencies that don't have a supported private route.
- Private paths for supported dependencies: Use eligible Search tiers, supported source configurations, approved outbound links, and the required execution environment. Configure each model and auxiliary dependency separately.
The private Foundry IQ tutorial covers a particular deployment, not every provider combination. Retain its documented trusted-service requirements; a trusted-service bypass isn't proof that an ingestion-time model call uses Private Link.
A network security perimeter governs supported public-endpoint access. It doesn't enable outbound shared private links on otherwise unsupported configurations.
Verify each configured path
Configure and validate supported paths before disabling public endpoints. For each dependency, record:
- The initiating runtime, destination endpoint, Search tier and hosting mode, region, API version, and execution mode.
- The identity, required roles, and token audience specified by the destination's setup owner.
- The configured network control, target subresource, provisioning state, and approval state where applicable.
- DNS resolution and reachability from the actual client for inbound connections.
- Applicable service diagnostics and execution history for the intended workload.
Use configuration, approval, routing, and diagnostic evidence together. A successful retrieve request, approved link, or completed indexer run alone doesn't establish that every dependency used a private path. Exercise ingestion, query models, remote sources, and auxiliary storage separately where your solution uses them.