Quickstart: Use agentic retrieval in the Azure portal

In this quickstart, you use agentic retrieval in the Azure portal to create a conversational search experience powered by documents indexed in Azure AI Search and large language models (LLMs) from Azure OpenAI in Foundry Models.

The portal guides you through the process of creating the following objects:

• A knowledge source that references a container in Azure Blob Storage. When you create a blob knowledge source, Azure AI Search automatically generates an index and other pipeline objects to ingest and enrich your content for agentic retrieval.

• A knowledge base that uses agentic retrieval to infer the underlying information need, plan and execute subqueries, and formulate a natural-language answer using the optional answer synthesis output mode.

Afterwards, you test the knowledge base by submitting a complex query that requires information from multiple documents and reviewing the synthesized answer.

Prerequisites

• An Azure account with an active subscription. Create an account for free.

• An Azure AI Search service in any region that provides agentic retrieval.

• An Azure Blob Storage account.

• A Microsoft Foundry project and resource. When you create a project, the resource is automatically created.

• For text-to-vector conversion, an embedding model deployed to your project. You can use any text-embedding model, such as text-embedding-3-large.

• For query planning and answer generation, an LLM deployed to your project. You can use any portal-supported LLM.

Supported LLMs

Although agentic retrieval programmatically supports several LLMs, the portal currently supports the following LLMs:

• gpt-4o
• gpt-4o-mini
• gpt-5
• gpt-5-mini
• gpt-5-nano

Configure access

Before you begin, make sure you have permissions to access content and operations. We recommend Microsoft Entra ID for authentication and role-based access for authorization. You must be an Owner or User Access Administrator to assign roles. If roles aren't feasible, use key-based authentication instead.

To configure access for this quickstart, select each of the following tabs.

Prepare sample data

This quickstart uses sample JSON documents from NASA's Earth at Night e-book, but you can also use your own files. The documents describe general science topics and images of Earth at night as observed from space.

To prepare the sample data for this quickstart:

1. Sign in to the Azure portal and select your Azure Blob Storage account.

2. From the left pane, select Data storage > Containers.

3. Create a container named earth-at-night-data.

4. Upload the sample JSON documents to the container.

Create a knowledge source

A knowledge source is a reusable reference to your source data. In this section, you create a blob knowledge source, which triggers the creation of a data source, skillset, index, and indexer to automate data indexing and enrichment. You review these objects in a later section.

You also configure a vectorizer, which uses your deployed embedding model to convert text into vectors and match documents based on semantic similarity. The vectorizer, vector fields, and vectors will be added to the auto-generated index.

To create the knowledge source for this quickstart:

1. Sign in to the Azure portal and select your search service.

2. From the left pane, select Knowledge retrieval > Knowledge sources.

3. Select Add knowledge source > Add knowledge source.

4. Enter earth-at-night-ks for the name.

5. Select Azure blob, and then select your subscription, storage account, and container with the sample data.

6. Select the Authenticate using managed identity checkbox. Leave the identity type as System-assigned.

7. Select Add vectorizer.

8. Select Microsoft Foundry for the kind, and then select your subscription, project, and embedding model deployment.

9. Select System managed identity for the authentication type.

10. Create the knowledge source.

    

Create a knowledge base

A knowledge base uses your knowledge source and deployed LLM to orchestrate agentic retrieval. When a user submits a complex query, the LLM generates subqueries that are sent simultaneously to your knowledge source. Azure AI Search then semantically ranks the results for relevance and combines the best results into a single, unified response.

The output mode determines how the knowledge base formulates answers. You can either use extractive data for verbatim content or answer synthesis for natural-language answer generation. By default, the portal uses answer synthesis.

To create the knowledge base for this quickstart:

1. From the left pane, select Knowledge retrieval > Knowledge bases.

2. Select Add knowledge base > Add knowledge base.

3. Enter earth-at-night-kb for the name.

4. Under Model deployment, select Add model deployment.

5. Select Foundry for the kind, and then select your subscription, project, and LLM deployment.

6. Select System assigned identity for the authentication type.

7. Save the model deployment.

8. Under Knowledge sources, select earth-at-night-ks.

9. Create the knowledge base.

   

Test agentic retrieval

The portal provides a chat playground where you can submit retrieve requests to the knowledge base, whose responses include references to your knowledge sources and debug information about the retrieval process.

To query the knowledge base:

1. Use the chat box to send the following query.

   Why do suburban belts display larger December brightening than urban cores even though absolute light levels are higher downtown? Why is the Phoenix nighttime street grid is so sharply visible from space, whereas large stretches of the interstate between midwestern cities remain comparatively dim?
   

2. Review the synthesized, citation-backed answer, which should be similar to the following example.

   The suburban belts exhibit larger December brightening compared to urban cores due to the increased use of decorative and festive lighting in residential areas, which are more prevalent in suburban regions. In contrast, urban cores, despite having higher absolute light levels, experience less seasonal variation in lighting. The Phoenix nighttime street grid is sharply visible from space because of its regular grid layout and the extensive use of street lighting, which creates a consistent and bright pattern. Conversely, large stretches of interstate highways between Midwestern cities are less illuminated, as they primarily serve as transit routes with minimal lighting infrastructure, resulting in comparatively dim visibility from space.
   

3. Select the debug icon to review the activity log, which should be similar to the following example.

   [
     {
       "type": "modelQueryPlanning",
       "id": 0,
       "inputTokens": 2081,
       "outputTokens": 128,
       "elapsedMs": 1577
     },
     {
       "type": "azureBlob",
       "id": 1,
       "knowledgeSourceName": "earth-at-night-ks",
       "queryTime": "2025-11-03T15:09:28.172Z",
       "count": 0,
       "elapsedMs": 731,
       "azureBlobArguments": {
         "search": "Why do suburban belts display larger December brightening than urban cores despite higher downtown light levels?"
       }
     },
     {
       "type": "azureBlob",
       "id": 2,
       "knowledgeSourceName": "earth-at-night-ks",
       "queryTime": "2025-11-03T15:09:28.669Z",
       "count": 3,
       "elapsedMs": 497,
       "azureBlobArguments": {
         "search": "Why is the Phoenix nighttime street grid sharply visible from space compared to dim interstates in the Midwest?"
       }
     },
     {
       "type": "semanticReranker",
       "id": 3,
       "inputTokens": 0
     },
     {
       "type": "modelAnswerSynthesis",
       "id": 4,
       "inputTokens": 3938,
       "outputTokens": 136,
       "elapsedMs": 1963
     }
   ]   
   

   The activity log offers insight into the steps taken during retrieval, including query planning and execution, semantic ranking, and answer synthesis. For more information, see Review the activity array.

Review the created objects

Azure AI Search automatically generates a data source, skillset, index, and indexer for each blob knowledge source. These objects form an end-to-end pipeline for data ingestion, enrichment, chunking, and vectorization. You can review these objects to learn how your data is processed for agentic retrieval.

To review the auto-generated objects:

1. From the left pane, select Search management.

2. Check the data source to verify the connection to your blob storage container.

3. Check the skillset to see how your content is chunked and vectorized using your embedding model.

4. Check the index to see how your content is indexed and exposed for retrieval, including which fields are searchable and filterable and which fields store vectors for similarity search.

5. Check the indexer for success or failure messages. Connection or quota errors appear here.

Clean up resources

When you work in your own subscription, it's a good idea to finish a project by determining whether you still need the resources you created. Resources that are left running can cost you money.

In the Azure portal, you can manage your Azure AI Search, Azure Blob Storage, and Foundry resources by selecting All resources or Resource groups from the left pane.

You can also delete the knowledge source and knowledge base on their respective portal pages. When you delete the knowledge source, the portal prompts you to delete the associated data source, skillset, index, and indexer.

Next step