07 - Add security filters for agent knowledge trimming

This module explains how to implement security filters to control knowledge access in your Azure Logic Apps agents, ensuring users only receive information they are authorized to view.

When finished with this module, you'll have gained the following knowledge:

• Security Filter Pattern: How to implement document-level authorization using user id filters in RAG-enabled Logic Apps

The Importance of Security in Enterprise AI Applications

In enterprise AI applications, ensuring that users can only access information they are authorized to view is crucial for maintaining data privacy, regulatory compliance, and organizational security. Security trimming prevents unauthorized access to sensitive documents and ensures that AI responses are contextually appropriate based on user permissions.

Security-Enhanced RAG Implementation

Implementing security filters in Azure Logic Apps RAG patterns requires a multi-layered approach that combines data indexing with security metadata and dynamic filtering during retrieval:

Phase 1: Secure Data Ingestion and Indexing

In this phase, Azure AI Search will index and vectorize your data:

• Parse and extract text content
• Generate embeddings for content
• Index documents with security fields in Azure AI Search, enabling efficient knowledge trimming during retrieval

Phase 2: Filtered Retrieval and Response Generation

When users interact with your agent, Logic Apps applies security filters before generating responses:

• Apply security filters to ensure only authorized content is retrieved
• Generate responses based only on content that the user is authorized to access

Phase 1 - Secure Data Ingestion and Indexing

note

Prerequisites for this phase are the following:

• You have access to an Azure Storage Account resource. For steps on setting this resource up, follow the guide here Create an Azure storage account.
• You have created a new container in your Azure Storage Account resource. For steps on creating a container, follow the guide here Manage blob containers using the Azure portal
• You have uploaded interns and employees folder to the container in your Azure Storage Account resource. The folders used in this module can be downloaded here: employees folder and interns folder in the container in your Azure Storage Account resource.
• You have access to an Azure OpenAI Service with a deployed text embeddings model. For more on creating this service, visit Explore Azure OpenAI in Azure AI Foundry.
• You have access to an Azure AI Search service. For more on creating this resource visit here Create an Azure AI Search service.
• You have configured an Azure AI Search service connection to the Azure Storage account using a managed identity : Connect to Azure Storage using a managed identity (Azure AI Search)

Step 1 - Index and Vectorize Data

1. In the Azure portal, open your Azure AI Search instance and click on Import Data (New).
2. Choose your data source and then select RAG.
3. On this page,
   1. Fill out the appropriate Subscription, Storage account and Blob container
   2. Leave Blob folder blank
   3. Parsing mode should be set to Text
   4. Check Authenticate using Managed Identity
   5. Select System-assigned or User-assigned based on how you configured the Azure AI Search service connection to the Azure Storage account
   6. Then click on Next
4. On the next page,
   1. Select AI Foundry Hub catalog models (Preview) for Kind
   2. Select the appropriate Subscription and Azure AI Foundry/Hub Project
   3. Select your embedding model under Model deployment
   4. Set API key as the Authentication type
   5. Check the acknowledgement
   6. Then click on Next
5. On the next page, keep all the default values and click on Next.
6. On the last page, rename your Objects name prefix to hr-index. Then press Create.

Step 2 - Add empty UserIds Field

1. On the left navigation bar of your Azure AI Search instance, click on Search management and then click on Indexes.
2. Click on hr-index.
3. Click on Fields and then click + Add field.
4. On the pop-up on the right-hand, fill in the following information:
   • Field Name: UserIds
   • Type: Collection(Edm.String)
   • Check Filterable
   • Press Save

Step 3 - Obtain document chunk ids

1. Start Visual Studio Code and select the Extensions tile.
2. Search for the Rest Client and Select Install.
3. Create a new file named get-documents.rest.
4. Paste in the following code in your file.

@endpoint = PUT-YOUR-SEARCH-SERVICE-ENDPOINT-HERE
@index = hr-index
@queryApiKey = PUT-YOUR-SEARCH-SERVICE-QUERY-API-KEY-HERE

POST https://{{endpoint}}/indexes/{{index}}/docs/search?api-version=2025-09-01
api-key: {{queryApiKey}}
Content-Type: application/json
    
{
   "search": "*",
   "select": "chunk_id,title",
   "top": 6,
   "skip": 0
}

You will need to replace the endpoint with your own. To do this,

1. On Azure Portal, open your Azure AI Search instance. Copy the link next to Url and paste it in get-documents.rest. The Url will look something like this: https://name-of-azure-search-service.search.windows.net.

You will need to replace the queryApiKey with your own. To do this,

1. On Azure Portal, open your Azure AI Search instance. On the left navigation bar, click on Settings and then click on Keys.
2. Under Manage query keys, copy the key and paste it in get-documents.rest.

Now press Send Request. You should have gotten a value array with the @search.score, chunk_id and title.

Step 4 - Populate User Ids for each document

note

This step demonstrates how to manually update documents with security metadata for testing purposes. In a production environment, you would typically incorporate this security metadata assignment into your automated data ingestion pipeline to ensure all documents are properly tagged with appropriate user access permissions during the indexing process.

1. Create a new file in Visual Studio Code named update-documents.rest
2. Paste in the following code in your file.

@endpoint = PUT-YOUR-SEARCH-SERVICE-ENDPOINT-HERE
@index = hr-index
@adminApiKey = PUT-YOUR-SEARCH-SERVICE-ADMIN-API-KEY-HERE

POST https://{{endpoint}}/indexes/{{index}}/docs/index?api-version=2025-09-01
api-key: {{adminApiKey}}
Content-Type: application/json

{
    "value": [
        {
            "@search.action": "mergeOrUpload",
            "chunk_id": "f9f29713558d_aHR0cHM6Ly93YXdzZHN0ZXN0c2E1MS5ibG9iLmNvcmUud2luZG93cy5uZXQvcmFnLXRlbXAtdGVzdC9pbnRlcm5zL2ludGVybl9qb2JfcmVzcG9uc2liaWxpdGllcy50eHQ1_pages_0",
            "title": "intern_job_responsibilities.txt",
            "UserIds": ["249f5761-6f30-4e27-a1db-df2399e13b3d"]
        },
        {
            "@search.action": "mergeOrUpload",
            "chunk_id": "f9f29713558d_aHR0cHM6Ly93YXdzZHN0ZXN0c2E1MS5ibG9iLmNvcmUud2luZG93cy5uZXQvcmFnLXRlbXAtdGVzdC9pbnRlcm5zL2ludGVybl9wZXJrcy50eHQ1_pages_0",
            "title": "intern_perks.txt",
            "UserIds": ["249f5761-6f30-4e27-a1db-df2399e13b3d"]
        },
        {
            "@search.action": "mergeOrUpload",
            "chunk_id": "f9f29713558d_aHR0cHM6Ly93YXdzZHN0ZXN0c2E1MS5ibG9iLmNvcmUud2luZG93cy5uZXQvcmFnLXRlbXAtdGVzdC9pbnRlcm5zL2ludGVybl9jb21wZW5zYXRpb24udHh00_pages_0",
            "title": "intern_compensation.txt",
            "UserIds": ["249f5761-6f30-4e27-a1db-df2399e13b3d"]
        },
        {
            "@search.action": "mergeOrUpload",
            "chunk_id": "f9f29713558d_aHR0cHM6Ly93YXdzZHN0ZXN0c2E1MS5ibG9iLmNvcmUud2luZG93cy5uZXQvcmFnLXRlbXAtdGVzdC9lbXBsb3llZXMvZW1wbG95ZWVfcGVya3MudHh00_pages_0",
            "title": "employee_perks.txt",
            "UserIds": ["cfb80024-5aab-49cb-af64-cc1df63aff24"]
        },
        {
            "@search.action": "mergeOrUpload",
            "chunk_id": "f9f29713558d_aHR0cHM6Ly93YXdzZHN0ZXN0c2E1MS5ibG9iLmNvcmUud2luZG93cy5uZXQvcmFnLXRlbXAtdGVzdC9lbXBsb3llZXMvZW1wbG95ZWVfam9iX3Jlc3BvbnNpYmlsaXRpZXMudHh00_pages_0",
            "title": "employee_job_responsibilities.txt",
            "UserIds": ["cfb80024-5aab-49cb-af64-cc1df63aff24"]
        },
        {
            "@search.action": "mergeOrUpload",
            "chunk_id": "f9f29713558d_aHR0cHM6Ly93YXdzZHN0ZXN0c2E1MS5ibG9iLmNvcmUud2luZG93cy5uZXQvcmFnLXRlbXAtdGVzdC9lbXBsb3llZXMvZW1wbG95ZWVfY29tcGVuc2F0aW9uLnR4dA2_pages_0",
            "title": "employee_compensation.txt",
            "UserIds": ["cfb80024-5aab-49cb-af64-cc1df63aff24"]
        }
    ]
}

Make sure to replace the endpoint with your own.

Update each chunk_id value in the value array with the actual chunk_id you obtained in Step 3. Make sure each chunk_id is being mapped to the correct document title.

Next, UserIds values must be updated with the actual Microsoft Entra object identifiers (GUIDs) of the users who should have access to view these documents. In this example:

• The first three documents (intern-related) use 249f5761-6f30-4e27-a1db-df2399e13b3d, which could represent the user id of a particular intern.
• The last three documents (employee-related) use cfb80024-5aab-49cb-af64-cc1df63aff24, which could represent the user id of a particular full-time employee.

You can assign multiple User Ids to each document by adding more GUIDs to the array if you want to grant access to multiple interns or full-time employees.

For testing purposes: You can use your own user Object ID for the employee documents (so you can access them) and use another user's Object ID for the intern documents (to test access restrictions).

To find your Object ID:

1. In the Azure Portal search box, search for your name
2. Your profile will appear under "Microsoft Entra ID"
3. Click on your profile and copy the "Object ID" value from the details page

Finally, you will need to replace the @adminApiKey with your own. To do this:

1. On Azure Portal, open your Azure AI Search instance. On the left navigation bar, click on Settings and then click on Keys.

2. Under Manage admin keys, copy the Primary admin key and paste it in update-documents.rest.

3. Now press Send Request.

Phase 2: Filtered Retrieval and Response Generation

note

Prerequisites for this module are the following:

• You have Easy Auth configured. For steps on setting up EasyAuth, follow the guide here Setting up Easy Auth.

Step 1 - Create the Logic Apps Workflow

1. In the Azure portal, open your Standard logic app resource. Add a new conversational agent workflow in the designer and name it Secure-Document-Retrieval.
2. Configure the agent with the following settings:

• Name: HR Agent
• System Instructions:

   You are an HR assistant that answers questions about job responsibilities, compensation and perks. Use the Document Search Tool to do a vector search of the user's question. Use only information outputted from this tool to answer the user's question. No other data or information should be used to answer the question.
  1. Click “Add an action” inside the agent loop.
  2. Select the Azure AI Search > Search vectors with natural language.
  3. Rename the tool: Document Search Tool
  4. Tool description: Query the knowledge base.
  5. Create an agent parameter:
     • Name: query
     • Type: String
  6. Add a new connection for the Search vectors with natural language action by setting the Azure AI Search Endpoint and Admin Key.
  7. Set the following:
     • Index Name: hr-index
     • Search Text: Insert the agent parameter query
     • Number of nearest neighbors to return: 2
  8. Press Show All under Advanced Parameters and set the following:
     • Filter Condition: UserIds/any(u: u eq '@{currentRequest()['headers']['X-MS-CLIENT-PRINCIPAL-ID']}')
     • The Vector Fields To Search: "text_vector"
     • Fields To Select: "chunk"

Step 2 - Test the Workflow

On the designer toolbar, select Chat to open the chat client in the Azure portal. Select the chat client URL, which opens the chat client integrated with your logic app.

In the chat client interface, ask the following question: What are the perks?

Since you only have access to employee documents, the response is based on the employee documents.

In a new chat session, ask the following question: What is the intern compensation?

Since you only have access to employee documents, the agent is unable to provide a response.

Security Best Practices for RAG-Enabled Logic Apps

• Principle of Least Privilege: Grant users access only to the minimum information required for their role
• Document-Level Access Control: Use the UserIds field to specify exactly which users can access each document
• Role-Based Filtering: Implement different UserIds arrays for different organizational roles