Skip to content

Latest commit

 

History

History
191 lines (139 loc) · 9.28 KB

README.md

File metadata and controls

191 lines (139 loc) · 9.28 KB

Azure AI Agent Service-enterprise-demo

This sample demonstrates how to build a streaming enterprise agent using Azure AI Agent Service. The agent can answer questions in real time using local HR and company policy documents, integrate external context via Bing, using gpt-4o-2024-05-13.

YouTube

Features

This demo teaches developers how to:

  • Create or Reuse Agents Programmatically
    Demonstrates how to connect to an Azure AI Foundry hub, either create a new agent with customized instructions (using GPT-4o or any supported model), or reuse an existing agent.

  • Incorporate Vector Stores for Enterprise Data
    Automatically create or reuse a vector store containing local policy files (e.g. HR, PTO, etc.) for retrieval-augmented generation (RAG).
    Optional: If the default file search tool isn’t available, the notebook automatically attempts direct Azure AI Search integration via environment variables.

  • Integrate Server-Side Tools
    Illustrates adding tools—like Bing search, file search, and custom Python functions—into a single ToolSet, and how to intercept and log each tool call.

  • Extend Functionality with Azure Logic Apps
    Deploy a Logic App to enable the send_email functionality. This Logic App can be imported using the provided ARM template, and its HTTP endpoint can be integrated into the agent’s toolset.

  • Stream Real-Time Agent Responses
    Demonstrates a streaming approach for partial message updates from the agent, seamlessly handling partial tool invocation and chunked text output.

  • Build an Interactive Gradio UI
    Provides a Gradio-based chat interface that prompts the agent with user questions, displays partial tool calls and final results, and makes it easy to extend or adapt the UI.

gif demo

Use this demo as a reference for creating, deploying, and managing enterprise-scale AI agents with strong integration, data security, and real-time conversation capabilities.

Getting Started

Prerequisites

  • Python 3.9+
  • Visual Studio Code with the Python and Jupyter extensions
  • An Azure AI Foundry resource set up (see Azure AI Agent Service docs)

Installation & Setup

  1. Clone this repository:

    git clone https://github.com/Azure-Samples/azure-ai-agent-service-enterprise-demo.git

  2. Create a virtual environment (using venv as an example):

    python -m venv .venv

  3. Activate your virtual environment:

    • Windows: .venv\Scripts\activate
    • macOS/Linux: source .venv/bin/activate

  4. Install the required dependencies:

    pip install -r requirements.txt

  5. Create a .env file at the root of this folder to store secret keys and settings (e.g., the connection string and optional model name). You can copy the provided .env.example file:

    • Windows (PowerShell):

      Copy-Item -Path .env.example -Destination .env

    • macOS/Linux:

      cp .env.example .env

    Then, open the .env file and update it with your configuration details.

    • Add your Azure AI Foundry connection string:

      PROJECT_CONNECTION_STRING="<HostName>;<AzureSubscriptionId>;<ResourceGroup>;<ProjectName>"

    • Specify the compatible model you want to use (e.g. GPT-4o):

      MODEL_NAME="YOUR_MODEL_NAME"

    • (Optional) Grounding with Bing

      You can add real-time web data to your agent via Grounding with Bing Search. For details on how to create a Bing search resource, link it with your Azure AI Agent, and meet display requirements, see Grounding with Bing Search.

      BING_CONNECTION_NAME="YOUR_CONNECTION_NAME"

      In this sample, the code automatically tries to discover an .env variable named BING_CONNECTION_NAME. If available, you’ll see a console message like bing > connected. Otherwise, it gracefully proceeds without Bing.

    • (Optional) OpenWeather API keys to enable fetch_weather tool:

      OPENWEATHER_GEO_API_KEY="YOUR_OPENWEATHER_GEOCODING_API_KEY"
OPENWEATHER_ONE_API_KEY="YOUR_OPENWEATHER_ONE_CALL_API_KEY"

      If you leave these blank, the weather function will simply return an error or remain disabled.

      Tip: If you don’t plan to use Bing grounding or OpenWeather, you can skip setting up those resources. The demo will still work with your local documents and core agent features.

    Make sure that .env is listed in your .gitignore. Never commit your credentials to source control!

  6. Open the project folder in Visual Studio Code:

    • Select your Python interpreter:
      1. Press Ctrl+Shift+P (or Cmd+Shift+P on Mac)
      2. Choose Python: Select Interpreter and select the venv environment.

Quickstart

  1. Run Jupyter Notebook:

    • Open the enterprise-streaming-agent.ipynb, in VS Code.
    • Step through the cells to:
      1. Connect to Azure AI Foundry and create or reuse an agent.
      2. Optionally upload local HR/policy files to a vector store.
      3. Add Bing integration, local file search, and custom Python functions (weather, stock lookup, etc.) to the agent’s ToolSet.
      4. If no FileSearchTool is detected, the code uses AZURE_SEARCH_CONNECTION_NAME and AZURE_SEARCH_INDEX_NAME from your .env file to add the search tool.
      5. Launch a Gradio UI that streams real-time queries and partial responses.

  2. Try the Demo Chat:

    • When you run the notebook, a local Gradio instance should launch in your cell output. You can click the localhost link to open the chat UI.
    • Ask the agent questions like:
      • “What’s my company’s remote work policy?”
      • “Fetch the weather forecast for Seattle tomorrow.” (Requires valid OpenWeather keys)
      • “How is MSFT stock price trending today?”
      • “Send an email summary of the HR policy.” (Triggers the Logic App if configured)

Deploying the Send Email Logic App

The sample includes a Logic App ARM template (send_email_logic_app.template.json) that you can deploy to enable the send_email functionality.

Steps to Deploy:

  1. The template defines a simple logic app that triggers on an HTTP request. It expects a JSON payload with recipient, subject, and body fields.
  2. Deploy the template using the Azure CLI or the Azure Portal.
    • Azure CLI: 
      az deployment group create \
  --resource-group <your-resource-group> \
  --template-file send_email_logic_app.template.json \
  --parameters logicAppName=send_email_logic_app
    • Azure Portal:
      • Go to the Azure Portal and create a new Logic App.
      • Choose the Blank Logic App template.
      • In the designer, add an HTTP trigger and an Office 365 Send an email action.
      • Save the Logic App and copy the HTTP endpoint URL.
  3. Once deployed, copy the HTTP trigger URL from the Logic App’s trigger.
  4. Uncomment and set the LOGIC_APP_SEND_EMAIL_URL variable with you Logic App URL: 
    LOGIC_APP_SEND_EMAIL_URL="https://<your-logic-app-endpoint>"

Azure AI Search Integration

This demo supports two approaches for enterprise document retrieval:

Default Vector Store

By default, the FileSearchTool automatically creates a vector store using Azure AI Search in standard agent setup, providing:

  • Automatic document chunking and embedding
  • Vector + keyword hybrid search
  • Zero additional configuration needed

Direct Azure AI Search Integration

For scenarios requiring direct control over an existing search index, update these environment variables to your .env:

#AZURE_SEARCH_CONNECTION_NAME="YOUR_AZURE_SEARCH_CONNECTION_NAME"
#AZURE_SEARCH_INDEX_NAME="YOUR_AZURE_SEARCH_INDEX_NAME"

Resources

Known Issues

Please review our Known Issues for current bugs and workarounds before reporting new problems.

Acknowledgments

  • Gradio
    This project uses Gradio under the Apache License 2.0. No modifications to Gradio’s source code are distributed in this repository.