Add Azure AI Search as a Copilot Studio Knowledge Source

Microsoft Copilot Studio can use an Azure AI Search vector index as a knowledge source for grounded answers. This is useful when an organization already has a searchable document collection and wants an agent to retrieve relevant passages and return links to the underlying material.

The connection does not make a poor index accurate or permission-aware by itself. Prepare the index, choose the authentication model, expose useful citation fields, and test both relevant and unauthorized queries before publishing the agent.

Prerequisites

• An Azure subscription and an Azure AI Search service.
• A Copilot Studio environment and an agent that supports the standard knowledge-source experience.
• A vector index in Azure AI Search. Copilot Studio supports indexes created with Azure AI Search integrated vectorization.
• Permission to create the Power Platform connection and read the selected search index.
• A supported authentication method and, for private networking, configured Power Platform virtual-network support.

Confirm regional availability, licensing, and capacity for both products. Search indexing, semantic ranking, embedding generation, and Copilot Studio usage can have separate costs.

Prepare the Azure AI Search index

Create or choose an index that contains the fields the agent needs for retrieval and citations. A typical document index includes:

• a unique document or chunk key;
• searchable text or chunk content;
• a vector field produced by the chosen embedding model;
• document title and other useful metadata;
• a complete URL for the source document; and
• optional filterable fields such as product, region, language, or access group.

Azure AI Search's Import and vectorize data workflow can chunk content, create embeddings, and configure an integrated vectorizer. With integrated vectorization, the search service can use the corresponding vectorizer for incoming queries instead of requiring custom prompt-embedding code.

If you want semantic ranking, enable and configure the semantic ranker in Azure AI Search before connecting the index. It is a separate feature and does not fix missing content, poor chunk boundaries, or incorrect metadata.

Add the knowledge source in Copilot Studio

1. Open the agent in Copilot Studio.
2. Select Add knowledge from the Overview or Knowledge page. You can also open the properties of a Generative Answers node when configuring knowledge for that node.
3. In Add knowledge, select Featured, then choose Azure AI Search.
4. Select Create new connection.
5. Choose one of the authentication types offered in your environment:
   • Microsoft Entra ID Integrated
   • Service principal (Microsoft Entra ID application)
   • Client Certificate Auth
   • Access Key, where supported
6. Enter the values required by that method. For example, an Entra integrated connection uses the search endpoint and signed-in identity; a key connection asks for the Azure AI Search endpoint and admin key.
7. Select Create and wait for the green confirmation.
8. Select Next, choose the vector index, and select Add to agent. Microsoft currently documents one Azure AI Search vector index per added knowledge-source connection.

The knowledge table shows In progress while Copilot Studio processes index metadata. Test only after the status changes to Ready.

Use the formal data connection

Add Azure AI Search through Data sources > Azure AI Search or the featured knowledge-source flow. Do not construct an unsupported connection by manually combining an endpoint and secret in another connector or custom field.

Microsoft warns that an unsupported connection can leave a failed environment-level data connection and prevent the Azure AI Search dialog from loading for other agents. Its current recovery guidance is to reset the agent's external access or recreate the affected agent, then add the source through the formal flow using Microsoft Entra ID authentication.

Although Access Key is displayed as an authentication option in the documented connection flow, prefer identity-based authentication when it meets the scenario. It avoids distributing a broad admin secret and can be governed with Azure role assignments. If a key is required, store and rotate it according to organizational policy and never paste it into an agent prompt or description.

Configure authentication and authorization

Connection authentication determines how Copilot Studio reaches Azure AI Search. It does not automatically enforce document-level permissions inside a shared index.

• Entra integrated: uses the signed-in connection identity. Confirm the assigned Azure AI Search data-plane roles.
• Service principal: uses an application identity. Grant only the roles required to query the intended service and index.
• Client certificate: avoids a plain client secret but requires certificate storage, rotation, and expiry monitoring.
• Access key: grants the capabilities associated with the key and should be treated as a sensitive secret.

If different agent users should see different documents, design and test an authorization pattern that filters results for the current user or separates indexes by security boundary. A single service connection can otherwise retrieve content that the agent user could not open directly.

Return usable citations

Include a field containing the complete URL of the source document. Copilot Studio recognizes metadata_storage_path as a citation field when it exists. If that field is absent, it can use another field containing a full URL.

The URL should open the authoritative source, not a temporary indexing file or inaccessible storage path. Users also need permission to open the cited destination. A citation that points to a restricted page may appear in an answer but fail when selected.

Store a readable title as well as the URL. During testing, verify that:

• the citation belongs to the passage used in the answer;
• the link opens the correct document and version;
• authorized users can open it;
• unauthorized users cannot retrieve or infer restricted content; and
• deleted or superseded documents stop appearing after the expected indexing delay.

Test retrieval before publishing

Use the agent's test panel and inspect the citations or activity details for each answer. Build a small evaluation set containing:

• questions with one clear answer;
• questions requiring information from several chunks;
• terms with common synonyms or abbreviations;
• questions whose answer is not in the index;
• contradictory or versioned documents;
• queries from authorized and unauthorized test users; and
• prompt-injection text embedded in a document.

A good result should cite the intended source and decline when evidence is missing. If retrieval is weak, improve the index: chunk size and overlap, searchable fields, vectorizer, semantic configuration, titles, filters, or source content. Prompt wording cannot compensate for a document that was never indexed correctly.

Use Azure AI Search behind a private endpoint

Copilot Studio supports Azure AI Search indexes configured for a virtual network and private endpoint. This requires both sides of the path to be configured: create the Azure AI Search private endpoint and set up virtual-network support for the Power Platform environment.

Validate DNS resolution, subnet delegation, network policies, tenant and region requirements, and the identity used by the connection. Do not disable the public endpoint until the private path works in a test environment and you have a recovery plan.

Troubleshooting

The connection dialog flashes and then fails

A broken environment-level connection may have been saved. Follow Microsoft's current recovery guidance rather than repeatedly entering keys. Because the connection can affect all agents in the environment, coordinate the fix with the Power Platform administrator.

No indexes appear

Confirm the endpoint, authentication identity, Azure role assignments, supported region, and that the index is a compatible vector index. Verify the same identity can query the service outside Copilot Studio using an approved administrative test.

The knowledge source stays In progress

Wait for metadata processing, then review service health and connection status. If it remains stuck, check recent index changes and environment diagnostics before deleting a production agent.

Answers have no useful citation

Add a full source URL to metadata_storage_path or another index field, then reindex. Confirm the URL is included in retrieved documents and is accessible to the test user.

Answers cite the wrong document

Review chunk boundaries, duplicate content, version metadata, semantic configuration, and the retrieved result set. Remove obsolete documents or add a reliable “current version” filter instead of asking the language model to guess which copy is authoritative.

Deployment checklist

• Use the supported Azure AI Search connection flow.
• Prefer a least-privilege Entra identity where practical.
• Keep secrets out of prompts, topics, and source fields.
• Provide stable, permission-appropriate citation URLs.
• Test negative, stale-data, authorization, and prompt-injection cases.
• Monitor indexing failures, key or certificate rotation, query quality, and costs.
• Publish first to a limited audience and keep a rollback path.

Microsoft updates this integration frequently. Check the official Copilot Studio Azure AI Search documentation for current authentication, networking, and recovery requirements before changing a production environment.