Поделиться через

Azure AI Projects client library for Python - version 1.0.0b6

  • Статья

Use the AI Projects client library (in preview) to:

  • Enumerate connections in your Azure AI Foundry project and get connection properties. For example, get the inference endpoint URL and credentials associated with your Azure OpenAI connection.
  • Get an authenticated Inference client to do chat completions, for the default Azure OpenAI or AI Services connections in your Azure AI Foundry project. Supports the AzureOpenAI client from the openai package, or clients from the azure-ai-inference package.
  • Develop Agents using the Azure AI Agent Service, leveraging an extensive ecosystem of models, tools, and capabilities from OpenAI, Microsoft, and other LLM providers. The Azure AI Agent Service enables the building of Agents for a wide range of generative AI use cases. The package is currently in preview.
  • Run Evaluations to assess the performance of generative AI applications using various evaluators and metrics. It includes built-in evaluators for quality, risk, and safety, and allows custom evaluators for specific needs.
  • Enable OpenTelemetry tracing.

Product documentation | Samples | API reference documentation | Package (PyPI) | SDK source code | AI Starter Template

Reporting issues

To report an issue with the client library, or request additional features, please open a GitHub issue here. Mention the package name "azure-ai-projects" in the title or content.

Table of contents

Getting started

Prerequisite

  • Python 3.8 or later.
  • An Azure subscription.
  • A project in Azure AI Foundry.
  • The project connection string. It can be found in your Azure AI Foundry project overview page, under "Project details". Below we will assume the environment variable PROJECT_CONNECTION_STRING was defined to hold this value.
  • Entra ID is needed to authenticate the client. Your application needs an object that implements the TokenCredential interface. Code samples here use DefaultAzureCredential. To get that working, you will need:
    • The Contributor role. Role assigned can be done via the "Access Control (IAM)" tab of your Azure AI Project resource in the Azure portal.
    • Azure CLI installed.
    • You are logged into your Azure account by running az login.
    • Note that if you have multiple Azure subscriptions, the subscription that contains your Azure AI Project resource must be your default subscription. Run az account list --output table to list all your subscription and see which one is the default. Run az account set --subscription "Your Subscription ID or Name" to change your default subscription.

Install the package

pip install azure-ai-projects

Key concepts

Create and authenticate the client

The class factory method from_connection_string is used to construct the client. To construct a synchronous client:

import os
from azure.ai.projects import AIProjectClient
from azure.identity import DefaultAzureCredential

project_client = AIProjectClient.from_connection_string(
    credential=DefaultAzureCredential(),
    conn_str=os.environ["PROJECT_CONNECTION_STRING"],
)

To construct an asynchronous client, Install the additional package aiohttp:

pip install aiohttp

and update the code above to import asyncio, and import AIProjectClient from the azure.ai.projects.aio namespace:

import os
import asyncio
from azure.ai.projects.aio import AIProjectClient
from azure.core.credentials import AzureKeyCredential

project_client = AIProjectClient.from_connection_string(
    credential=DefaultAzureCredential(),
    conn_str=os.environ["PROJECT_CONNECTION_STRING"],
)

Examples

Enumerate connections

Your Azure AI Foundry project has a "Management center". When you enter it, you will see a tab named "Connected resources" under your project. The .connections operations on the client allow you to enumerate the connections and get connection properties. Connection properties include the resource URL and authentication credentials, among other things.

Below are code examples of the connection operations. Full samples can be found under the "connetions" folder in the package samples.

Get properties of all connections

To list the properties of all the connections in the Azure AI Foundry project:

connections = project_client.connections.list()
for connection in connections:
    print(connection)

Get properties of all connections of a particular type

To list the properties of connections of a certain type (here Azure OpenAI):

connections = project_client.connections.list(
    connection_type=ConnectionType.AZURE_OPEN_AI,
)
for connection in connections:
    print(connection)

Get properties of a default connection

To get the properties of the default connection of a certain type (here Azure OpenAI), with its authentication credentials:

connection = project_client.connections.get_default(
    connection_type=ConnectionType.AZURE_OPEN_AI,
    include_credentials=True,  # Optional. Defaults to "False".
)
print(connection)

If the call was made with include_credentials=True, depending on the value of connection.authentication_type, either connection.key or connection.token_credential will be populated. Otherwise both will be None.

Get properties of a connection by its connection name

To get the connection properties of a connection named connection_name:

connection = project_client.connections.get(
    connection_name=connection_name,
    include_credentials=True  # Optional. Defaults to "False"
)
print(connection)

Get an authenticated ChatCompletionsClient

Your Azure AI Foundry project may have one or more AI models deployed that support chat completions. These could be OpenAI models, Microsoft models, or models from other providers. Use the code below to get an already authenticated ChatCompletionsClient from the azure-ai-inference package, and execute a chat completions call.

First, install the package:

pip install azure-ai-inference

Then run this code (replace "gpt-4o" with your model deployment name):

inference_client = project_client.inference.get_chat_completions_client()

response = inference_client.complete(
    model="gpt-4o", # Model deployment name
    messages=[UserMessage(content="How many feet are in a mile?")]
)

print(response.choices[0].message.content)

See the "inference" folder in the package samples for additional samples, including getting an authenticated EmbeddingsClient and ImageEmbeddingsClient.

Get an authenticated AzureOpenAI client

Your Azure AI Foundry project may have one or more OpenAI models deployed that support chat completions. Use the code below to get an already authenticated AzureOpenAI from the openai package, and execute a chat completions call.

First, install the package:

pip install openai

Then run the code below. Replace gpt-4o with your model deployment name, and update the api_version value with one found in the "Data plane - inference" row in this table.

aoai_client = project_client.inference.get_azure_openai_client(api_version="2024-06-01")

response = aoai_client.chat.completions.create(
    model="gpt-4o", # Model deployment name
    messages=[
        {
            "role": "user",
            "content": "How many feet are in a mile?",
        },
    ],
)

print(response.choices[0].message.content)

See the "inference" folder in the package samples for additional samples.

Agents (Preview)

Agents in the Azure AI Projects client library are designed to facilitate various interactions and operations within your AI projects. They serve as the core components that manage and execute tasks, leveraging different tools and resources to achieve specific goals. The following steps outline the typical sequence for interacting with Agents. See the "agents" folder in the package samples for additional Agent samples.

Create Agent

Before creating an Agent, you need to set up Azure resources to deploy your model. Create a New Agent Quickstart details selecting and deploying your Agent Setup.

Here is an example of how to create an Agent:

agent = project_client.agents.create_agent(
    model=os.environ["MODEL_DEPLOYMENT_NAME"],
    name="my-assistant",
    instructions="You are helpful assistant",
)

To allow Agents to access your resources or custom functions, you need tools. You can pass tools to create_agent by either toolset or combination of tools and tool_resources.

Here is an example of toolset:

functions = FunctionTool(user_functions)
code_interpreter = CodeInterpreterTool()

toolset = ToolSet()
toolset.add(functions)
toolset.add(code_interpreter)

agent = project_client.agents.create_agent(
    model=os.environ["MODEL_DEPLOYMENT_NAME"],
    name="my-assistant",
    instructions="You are a helpful assistant",
    toolset=toolset,
)

Also notices that if you use asynchronous client, you use AsyncToolSet instead. Additional information related to AsyncFunctionTool be discussed in the later sections.

Here is an example to use tools and tool_resources:

file_search_tool = FileSearchTool(vector_store_ids=[vector_store.id])

# Notices that FileSearchTool as tool and tool_resources must be added or the assistant unable to search the file
agent = project_client.agents.create_agent(
    model=os.environ["MODEL_DEPLOYMENT_NAME"],
    name="my-assistant",
    instructions="You are helpful assistant",
    tools=file_search_tool.definitions,
    tool_resources=file_search_tool.resources,
)

In the following sections, we show you sample code in either toolset or combination of tools and tool_resources.

To perform file search by an Agent, we first need to upload a file, create a vector store, and associate the file to the vector store. Here is an example:

file = project_client.agents.upload_file_and_poll(file_path="product_info_1.md", purpose="assistants")
print(f"Uploaded file, file ID: {file.id}")

vector_store = project_client.agents.create_vector_store_and_poll(file_ids=[file.id], name="my_vectorstore")
print(f"Created vector store, vector store ID: {vector_store.id}")

# Create file search tool with resources followed by creating agent
file_search = FileSearchTool(vector_store_ids=[vector_store.id])

agent = project_client.agents.create_agent(
    model=os.environ["MODEL_DEPLOYMENT_NAME"],
    name="my-assistant",
    instructions="Hello, you are helpful assistant and can search information from uploaded files",
    tools=file_search.definitions,
    tool_resources=file_search.resources,
)

We can upload file to Azure as it is shown in the example, or use the existing Azure blob storage. In the code below we demonstrate how this can be achieved. First we upload file to azure and create VectorStoreDataSource, which then is used to create vector store. This vector store is then given to the FileSearchTool constructor.

# We will upload the local file to Azure and will use it for vector store creation.
_, asset_uri = project_client.upload_file("./product_info_1.md")

# Create a vector store with no file and wait for it to be processed
ds = VectorStoreDataSource(asset_identifier=asset_uri, asset_type=VectorStoreDataSourceAssetType.URI_ASSET)
vector_store = project_client.agents.create_vector_store_and_poll(data_sources=[ds], name="sample_vector_store")
print(f"Created vector store, vector store ID: {vector_store.id}")

# Create a file search tool
file_search_tool = FileSearchTool(vector_store_ids=[vector_store.id])

# Notices that FileSearchTool as tool and tool_resources must be added or the assistant unable to search the file
agent = project_client.agents.create_agent(
    model=os.environ["MODEL_DEPLOYMENT_NAME"],
    name="my-assistant",
    instructions="You are helpful assistant",
    tools=file_search_tool.definitions,
    tool_resources=file_search_tool.resources,
)

We also can attach files to the existing vector store. In the code snippet below, we first create an empty vector store and add file to it.

# Create a vector store with no file and wait for it to be processed
vector_store = project_client.agents.create_vector_store_and_poll(data_sources=[], name="sample_vector_store")
print(f"Created vector store, vector store ID: {vector_store.id}")

ds = VectorStoreDataSource(asset_identifier=asset_uri, asset_type=VectorStoreDataSourceAssetType.URI_ASSET)
# Add the file to the vector store or you can supply data sources in the vector store creation
vector_store_file_batch = project_client.agents.create_vector_store_file_batch_and_poll(
    vector_store_id=vector_store.id, data_sources=[ds]
)
print(f"Created vector store file batch, vector store file batch ID: {vector_store_file_batch.id}")

# Create a file search tool
file_search_tool = FileSearchTool(vector_store_ids=[vector_store.id])

Create Agent with Code Interpreter

Here is an example to upload a file and use it for code interpreter by an Agent:

file = project_client.agents.upload_file_and_poll(
    file_path="nifty_500_quarterly_results.csv", purpose=FilePurpose.AGENTS
)
print(f"Uploaded file, file ID: {file.id}")

code_interpreter = CodeInterpreterTool(file_ids=[file.id])

# Create agent with code interpreter tool and tools_resources
agent = project_client.agents.create_agent(
    model=os.environ["MODEL_DEPLOYMENT_NAME"],
    name="my-assistant",
    instructions="You are helpful assistant",
    tools=code_interpreter.definitions,
    tool_resources=code_interpreter.resources,
)

Create Agent with Bing Grounding

To enable your Agent to perform search through Bing search API, you use BingGroundingTool along with a connection.

Here is an example:

bing_connection = project_client.connections.get(connection_name=os.environ["BING_CONNECTION_NAME"])
conn_id = bing_connection.id

print(conn_id)

# Initialize agent bing tool and add the connection id
bing = BingGroundingTool(connection_id=conn_id)

# Create agent with the bing tool and process assistant run
with project_client:
    agent = project_client.agents.create_agent(
        model=os.environ["MODEL_DEPLOYMENT_NAME"],
        name="my-assistant",
        instructions="You are a helpful assistant",
        tools=bing.definitions,
        headers={"x-ms-enable-preview": "true"},
    )

Azure AI Search is an enterprise search system for high-performance applications. It integrates with Azure OpenAI Service and Azure Machine Learning, offering advanced search technologies like vector search and full-text search. Ideal for knowledge base insights, information discovery, and automation. Creating an Agent with Azure AI Search requires an existing Azure AI Search Index. For more information and setup guides, see Azure AI Search Tool Guide.

Here is an example to integrate Azure AI Search:

conn_list = project_client.connections.list()
conn_id = ""
for conn in conn_list:
    if conn.connection_type == ConnectionType.AZURE_AI_SEARCH:
        conn_id = conn.id
        break

print(conn_id)

# Initialize agent AI search tool and add the search index connection id
ai_search = AzureAISearchTool(index_connection_id=conn_id, index_name="myindexname")

# Create agent with AI search tool and process assistant run
with project_client:
    agent = project_client.agents.create_agent(
        model=os.environ["MODEL_DEPLOYMENT_NAME"],
        name="my-assistant",
        instructions="You are a helpful assistant",
        tools=ai_search.definitions,
        tool_resources=ai_search.resources,
    )

Create Agent with Function Call

You can enhance your Agents by defining callback functions as function tools. These can be provided to create_agent via either the toolset parameter or the combination of tools and tool_resources. Here are the distinctions:

  • toolset: When using the toolset parameter, you provide not only the function definitions and descriptions but also their implementations. The SDK will execute these functions within create_and_run_process or streaming . These functions will be invoked based on their definitions.
  • tools and tool_resources: When using the tools and tool_resources parameters, only the function definitions and descriptions are provided to create_agent, without the implementations. The Run or event handler of stream will raise a requires_action status based on the function definitions. Your code must handle this status and call the appropriate functions.

For more details about calling functions by code, refer to sample_agents_stream_eventhandler_with_functions.py and sample_agents_functions.py.

For more details about requirements and specification of functions, refer to Function Tool Specifications

Here is an example to use user functions in toolset:

functions = FunctionTool(user_functions)
toolset = ToolSet()
toolset.add(functions)

agent = project_client.agents.create_agent(
    model=os.environ["MODEL_DEPLOYMENT_NAME"],
    name="my-assistant",
    instructions="You are a helpful assistant",
    toolset=toolset,
)

For asynchronous functions, you must import AIProjectClient from azure.ai.projects.aio and use AsyncFunctionTool. Here is an example using asynchronous user functions:

from azure.ai.projects.aio import AIProjectClient

functions = AsyncFunctionTool(user_async_functions)

toolset = AsyncToolSet()
toolset.add(functions)

agent = await project_client.agents.create_agent(
    model=os.environ["MODEL_DEPLOYMENT_NAME"],
    name="my-assistant",
    instructions="You are a helpful assistant",
    toolset=toolset,
)

Create Agent With Azure Function Call

The agent can handle Azure Function calls on the service side and return the result of the call. To use the function, we need to create the AzureFunctionTool, which contains the input and output queues of Azure function and the description of input parameters. Please note that in the prompt we are asking the model to invoke queue when the specific question ("What would foo say?") is being asked. See below for the instructions on function deployment.

azure_function_tool = AzureFunctionTool(
    name="foo",
    description="Get answers from the foo bot.",
    parameters={
        "type": "object",
        "properties": {
            "query": {"type": "string", "description": "The question to ask."},
            "outputqueueuri": {"type": "string", "description": "The full output queue uri."},
        },
    },
    input_queue=AzureFunctionStorageQueue(
        queue_name="azure-function-foo-input",
        storage_service_endpoint=storage_service_endpoint,
    ),
    output_queue=AzureFunctionStorageQueue(
        queue_name="azure-function-tool-output",
        storage_service_endpoint=storage_service_endpoint,
    ),
)

agent = project_client.agents.create_agent(
    model="gpt-4",
    name="azure-function-agent-foo",
    instructions=f"You are a helpful support agent. Use the provided function any time the prompt contains the string 'What would foo say?'. When you invoke the function, ALWAYS specify the output queue uri parameter as '{storage_service_endpoint}/azure-function-tool-output'. Always responds with \"Foo says\" and then the response from the tool.",
    tools=azure_function_tool.definitions,
)
print(f"Created agent, agent ID: {agent.id}")

To make a function call we need to create and deploy the Azure function. In the code snippet below, we have an example of function on python which can be used by the code above.

import azure.functions as func
import json

from urllib.parse import urlparse
from azure.identity import DefaultAzureCredential
from azure.storage.queue import (
    QueueClient,
    BinaryBase64EncodePolicy,
    BinaryBase64DecodePolicy
)

app = func.FunctionApp()


@app.function_name(name="Foo")
@app.queue_trigger(
    arg_name="arguments",
    queue_name="azure-function-foo-input",
    connection="AzureWebJobsStorage")
def foo(arguments: func.QueueMessage) -> None:
    """
    The function, answering question.

    :param arguments: The arguments, containing json serialized request.
    """
    parsed_args = json.loads(arguments.get_body().decode('utf-8'))
    queue_url = urlparse(parsed_args['outputqueueuri'])

    queue_client = QueueClient(
        f"{queue_url.scheme}://{queue_url.netloc}",
        queue_name=queue_url.path[1:],
        credential=DefaultAzureCredential(),
        message_encode_policy=BinaryBase64EncodePolicy(),
        message_decode_policy=BinaryBase64DecodePolicy()
    )

    response = {
        "Value": "Bar",
        "CorrelationId": parsed_args['CorrelationId']
    }
    queue_client.send_message(json.dumps(response).encode('utf-8'))

In this code we define function input and output class: Arguments and Response respectively. These two data classes will be serialized in JSON. It is important that these both contain field CorrelationId, which is the same between input and output.

In our example the function will be stored in the storage account, created with the AI hub. For that we need to allow key access to that storage. In Azure portal go to Storage account > Settings > Configuration and set "Allow storage account key access" to Enabled. If it is not done, the error will be displayed "The remote server returned an error: (403) Forbidden." Before creation of the function we will need to get the python version using command python --version. We recommend to use python version 3.11 or above. We will need only two major digits in the next command, which deploys function and installs azure-cli:

pip install -U azure-cli
az login
az functionapp create --resource-group your-resource-group --consumption-plan-location region --runtime python --runtime-version 3.11 --functions-version 4 --name function_name --os-type linux --storage-account storage_account_already_present_in_resource_group --app-insights existing_or_new_application_insights_name

This function writes data to the output queue and hence needs to be authenticated to Azure, so we will need to assign the function system identity and provide it Storage Queue Data Contributor. To do that in Azure portal select the function, located in your-resource-group resource group and in Settings>Identity, switch it on and click Save. After that assign the Storage Queue Data Contributor permission on storage account used by our function (storage_account_already_present_in_resource_group in the script above) for just assigned System Managed identity. Note: in python we need to provide the explicit queue connection. It is defined in the line connection="AzureWebJobsStorage".. AzureWebJobsStorage is a setting, which can be viewed in function_name Settings>Environment variables>AzureWebJobsStorage and will look like DefaultEndpointsProtocol=https;EndpointSuffix=core.windows.net;AccountName=storage_account_already_present_in_resource_group;AccountKey=xxxxx. Another option is to provide the Managed identity. In this case we will need to set connection to prefix, present in three environment variables. For example, connection=STORAGE_CONNECTION and the variables will be: STORAGE_CONNECTION__clientId (managed entity UUID), STORAGE_CONNECTION__credential ("managedidentity") and STORAGE_CONNECTION__queueServiceUri (the URI of storage account, for example https://storage_account_already_present_in_resource_group.queue.core.windows.net).

Now we will create the function itself. Install .NET and Core Tools and create the function project using next commands.

func init FunctionProj --worker-runtime python
cd FunctionProj
# Use the next line in cmd
echo.azure-identity>> requirements.txt
echo.azure-storage-queue>> requirements.txt
# Use the next line in PowerShell
echo "azure-identity" >> requirements.txt
echo "azure-storage-queue" >> requirements.txt

If you are working in virtual environment and the virtual environment folder is located in FunctionProj, make sure, the name of this folder is added to .funcignore as all the directory content will be uploaded to Azure.

Rename function_app.py to foo.py and replace the content with the code above. To deploy the function run the command from dotnet project folder:

func azure functionapp publish function_name

In the storage_account_already_present_in_resource_group select the Queue service and create two queues: azure-function-foo-input and azure-function-tool-output. Note that the same queues are used in our sample. To check that the function is working, place the next message into the azure-function-foo-input and replace storage_account_already_present_in_resource_group by the actual resource group name, or just copy the output queue address. Note in python outputqueueuri is written in all lower case letters.

{
  "outputqueueuri": "https://storage_account_already_present_in_resource_group.queue.core.windows.net/azure-function-tool-output",
  "CorrelationId": "42"
}

Next, we will monitor the output queue or the message. You should receive the next message.

{
  "Value": "Bar",
  "CorrelationId": "42"
}

Please note that the input CorrelationId is the same as output. Hint: Place multiple messages to input queue and keep second internet browser window with the output queue open and hit the refresh button on the portal user interface, so that you will not miss the message. If the message instead went to azure-function-foo-input-poison queue, the function completed with error, please check your setup. After we have tested the function and made sure it works, please make sure that the Azure AI Project have the following roles for the storage account: Storage Account Contributor, Storage Blob Data Contributor, Storage File Data Privileged Contributor, Storage Queue Data Contributor and Storage Table Data Contributor. Now the function is ready to be used by the agent.

Create Agent With Logic Apps

Logic Apps allow HTTP requests to trigger actions. For more information, refer to the guide Logic App Workflows for Function Calling.

Your Logic App must be in the same resource group as your Azure AI Project, shown in the Azure Portal. Agents SDK accesses Logic Apps through Workflow URLs, which are fetched and called as requests in functions.

Below is an example of how to create an Azure Logic App utility tool and register a function with it.


# Create the project client
project_client = AIProjectClient.from_connection_string(
    credential=DefaultAzureCredential(),
    conn_str=os.environ["PROJECT_CONNECTION_STRING"],
)

# Extract subscription and resource group from the project scope
subscription_id = project_client.scope["subscription_id"]
resource_group = project_client.scope["resource_group_name"]

# Logic App details
logic_app_name = "<LOGIC_APP_NAME>"
trigger_name = "<TRIGGER_NAME>"

# Create and initialize AzureLogicAppTool utility
logic_app_tool = AzureLogicAppTool(subscription_id, resource_group)
logic_app_tool.register_logic_app(logic_app_name, trigger_name)
print(f"Registered logic app '{logic_app_name}' with trigger '{trigger_name}'.")

# Create the specialized "send_email_via_logic_app" function for your agent tools
send_email_func = create_send_email_function(logic_app_tool, logic_app_name)

# Prepare the function tools for the agent
functions_to_use: Set = {
    fetch_current_datetime,
    send_email_func,  # This references the AzureLogicAppTool instance via closure
}

After this the functions can be incorporated normally into code using FunctionTool.

Create Agent With OpenAPI

OpenAPI specifications describe REST operations against a specific endpoint. Agents SDK can read an OpenAPI spec, create a function from it, and call that function against the REST endpoint without additional client-side execution.

Here is an example creating an OpenAPI tool (using anonymous authentication):


with open("./weather_openapi.json", "r") as f:
    openapi_weather = jsonref.loads(f.read())

with open("./countries.json", "r") as f:
    openapi_countries = jsonref.loads(f.read())

# Create Auth object for the OpenApiTool (note that connection or managed identity auth setup requires additional setup in Azure)
auth = OpenApiAnonymousAuthDetails()

# Initialize agent OpenApi tool using the read in OpenAPI spec
openapi_tool = OpenApiTool(
    name="get_weather", spec=openapi_weather, description="Retrieve weather information for a location", auth=auth
)
openapi_tool.add_definition(
    name="get_countries", spec=openapi_countries, description="Retrieve a list of countries", auth=auth
)

# Create agent with OpenApi tool and process assistant run
with project_client:
    agent = project_client.agents.create_agent(
        model=os.environ["MODEL_DEPLOYMENT_NAME"],
        name="my-assistant",
        instructions="You are a helpful assistant",
        tools=openapi_tool.definitions,
    )

Create Thread

For each session or conversation, a thread is required. Here is an example:

thread = project_client.agents.create_thread()

Create Thread with Tool Resource

In some scenarios, you might need to assign specific resources to individual threads. To achieve this, you provide the tool_resources argument to create_thread. In the following example, you create a vector store and upload a file, enable an Agent for file search using the tools argument, and then associate the file with the thread using the tool_resources argument.

file = project_client.agents.upload_file_and_poll(file_path="product_info_1.md", purpose="assistants")
print(f"Uploaded file, file ID: {file.id}")

vector_store = project_client.agents.create_vector_store_and_poll(file_ids=[file.id], name="my_vectorstore")
print(f"Created vector store, vector store ID: {vector_store.id}")

# Create file search tool with resources followed by creating agent
file_search = FileSearchTool(vector_store_ids=[vector_store.id])

agent = project_client.agents.create_agent(
    model=os.environ["MODEL_DEPLOYMENT_NAME"],
    name="my-assistant",
    instructions="Hello, you are helpful assistant and can search information from uploaded files",
    tools=file_search.definitions,
)

print(f"Created agent, ID: {agent.id}")

# Create thread with file resources.
# If the agent has multiple threads, only this thread can search this file.
thread = project_client.agents.create_thread(tool_resources=file_search.resources)

Create Message

To create a message for assistant to process, you pass user as role and a question as content:

message = project_client.agents.create_message(thread_id=thread.id, role="user", content="Hello, tell me a joke")

Create Message with File Search Attachment

To attach a file to a message for content searching, you use MessageAttachment and FileSearchTool:

attachment = MessageAttachment(file_id=file.id, tools=FileSearchTool().definitions)
message = project_client.agents.create_message(
    thread_id=thread.id, role="user", content="What feature does Smart Eyewear offer?", attachments=[attachment]
)

Create Message with Code Interpreter Attachment

To attach a file to a message for data analysis, use MessageAttachment and CodeInterpreterTool classes. You must pass CodeInterpreterTool as tools or toolset in create_agent call or the file attachment cannot be opened for code interpreter.

Here is an example to pass CodeInterpreterTool as tool:

# Notice that CodeInterpreter must be enabled in the agent creation,
# otherwise the agent will not be able to see the file attachment for code interpretation
agent = project_client.agents.create_agent(
    model=os.environ["MODEL_DEPLOYMENT_NAME"],
    name="my-assistant",
    instructions="You are helpful assistant",
    tools=CodeInterpreterTool().definitions,
)
print(f"Created agent, agent ID: {agent.id}")

thread = project_client.agents.create_thread()
print(f"Created thread, thread ID: {thread.id}")

# Create an attachment
attachment = MessageAttachment(file_id=file.id, tools=CodeInterpreterTool().definitions)

# Create a message
message = project_client.agents.create_message(
    thread_id=thread.id,
    role="user",
    content="Could you please create bar chart in TRANSPORTATION sector for the operating profit from the uploaded csv file and provide file to me?",
    attachments=[attachment],
)

Azure blob storage can be used as a message attachment. In this case, use VectorStoreDataSource as a data source:

# We will upload the local file to Azure and will use it for vector store creation.
_, asset_uri = project_client.upload_file("./product_info_1.md")
ds = VectorStoreDataSource(asset_identifier=asset_uri, asset_type=VectorStoreDataSourceAssetType.URI_ASSET)

# Create a message with the attachment
attachment = MessageAttachment(data_source=ds, tools=code_interpreter.definitions)
message = project_client.agents.create_message(
    thread_id=thread.id, role="user", content="What does the attachment say?", attachments=[attachment]
)

Create Run, Run_and_Process, or Stream

To process your message, you can use create_run, create_and_process_run, or create_stream.

create_run requests the Agent to process the message without polling for the result. If you are using function tools regardless as toolset or not, your code is responsible for polling for the result and acknowledging the status of Run. When the status is requires_action, your code is responsible for calling the function tools. For a code sample, visit sample_agents_functions.py.

Here is an example of create_run and poll until the run is completed:

run = project_client.agents.create_run(thread_id=thread.id, assistant_id=agent.id)

# Poll the run as long as run status is queued or in progress
while run.status in ["queued", "in_progress", "requires_action"]:
    # Wait for a second
    time.sleep(1)
    run = project_client.agents.get_run(thread_id=thread.id, run_id=run.id)

To have the SDK poll on your behalf and call function tools, use the create_and_process_run method. Note that function tools will only be invoked if they are provided as toolset during the create_agent call.

Here is an example:

run = project_client.agents.create_and_process_run(thread_id=thread.id, assistant_id=agent.id)

With streaming, polling need not be considered. If function tools are provided as toolset during the create_agent call, they will be invoked by the SDK.

Here is an example of streaming:

with project_client.agents.create_stream(thread_id=thread.id, assistant_id=agent.id) as stream:

    for event_type, event_data, _ in stream:

        if isinstance(event_data, MessageDeltaChunk):
            print(f"Text delta received: {event_data.text}")

        elif isinstance(event_data, ThreadMessage):
            print(f"ThreadMessage created. ID: {event_data.id}, Status: {event_data.status}")

        elif isinstance(event_data, ThreadRun):
            print(f"ThreadRun status: {event_data.status}")

        elif isinstance(event_data, RunStep):
            print(f"RunStep type: {event_data.type}, Status: {event_data.status}")

        elif event_type == AgentStreamEvent.ERROR:
            print(f"An error occurred. Data: {event_data}")

        elif event_type == AgentStreamEvent.DONE:
            print("Stream completed.")
            break

        else:
            print(f"Unhandled Event Type: {event_type}, Data: {event_data}")

In the code above, because an event_handler object is not passed to the create_stream function, the SDK will instantiate AgentEventHandler or AsyncAgentEventHandler as the default event handler and produce an iterable object with event_type and event_data. AgentEventHandler and AsyncAgentEventHandler are overridable. Here is an example:

# With AgentEventHandler[str], the return type for each event functions is optional string.
class MyEventHandler(AgentEventHandler[str]):

    def on_message_delta(self, delta: "MessageDeltaChunk") -> Optional[str]:
        return f"Text delta received: {delta.text}"

    def on_thread_message(self, message: "ThreadMessage") -> Optional[str]:
        return f"ThreadMessage created. ID: {message.id}, Status: {message.status}"

    def on_thread_run(self, run: "ThreadRun") -> Optional[str]:
        return f"ThreadRun status: {run.status}"

    def on_run_step(self, step: "RunStep") -> Optional[str]:
        return f"RunStep type: {step.type}, Status: {step.status}"

    def on_error(self, data: str) -> Optional[str]:
        return f"An error occurred. Data: {data}"

    def on_done(self) -> Optional[str]:
        return "Stream completed."

    def on_unhandled_event(self, event_type: str, event_data: Any) -> Optional[str]:
        return f"Unhandled Event Type: {event_type}, Data: {event_data}"

with project_client.agents.create_stream(
    thread_id=thread.id, assistant_id=agent.id, event_handler=MyEventHandler()
) as stream:
    for event_type, event_data, func_return in stream:
        print(f"Received data.")
        print(f"Streaming receive Event Type: {event_type}")
        print(f"Event Data: {str(event_data)[:100]}...")
        print(f"Event Function return: {func_return}\n")

As you can see, this SDK parses the events and produces various event types similar to OpenAI assistants. In your use case, you might not be interested in handling all these types and may decide to parse the events on your own. To achieve this, please refer to override base event handler.

Note: Multiple streaming processes may be chained behind the scenes.

When the SDK receives a `ThreadRun` event with the status `requires_action`, the next event will be `Done`, followed by termination. The SDK will submit the tool calls using the same event handler. The event handler will then chain the main stream with the tool stream.

Consequently, when you iterate over the streaming using a for loop similar to the example above, the for loop will receive events from the main stream followed by events from the tool stream.

Retrieve Message

To retrieve messages from agents, use the following example:

messages = project_client.agents.list_messages(thread_id=thread.id)

# The messages are following in the reverse order,
# we will iterate them and output only text contents.
for data_point in reversed(messages.data):
    last_message_content = data_point.content[-1]
    if isinstance(last_message_content, MessageTextContent):
        print(f"{data_point.role}: {last_message_content.text.value}")

In addition, messages and messages.data[] offer helper properties such as text_messages, image_contents, file_citation_annotations, and file_path_annotations to quickly retrieve content from one message or all messages.

Retrieve File

Files uploaded by Agents cannot be retrieved back. If your use case need to access the file content uploaded by the Agents, you are advised to keep an additional copy accessible by your application. However, files generated by Agents are retrievable by save_file or get_file_content.

Here is an example retrieving file ids from messages and save to the local drive:

messages = project_client.agents.list_messages(thread_id=thread.id)
print(f"Messages: {messages}")

for image_content in messages.image_contents:
    file_id = image_content.image_file.file_id
    print(f"Image File ID: {file_id}")
    file_name = f"{file_id}_image_file.png"
    project_client.agents.save_file(file_id=file_id, file_name=file_name)
    print(f"Saved image file to: {Path.cwd() / file_name}")

for file_path_annotation in messages.file_path_annotations:
    print(f"File Paths:")
    print(f"Type: {file_path_annotation.type}")
    print(f"Text: {file_path_annotation.text}")
    print(f"File ID: {file_path_annotation.file_path.file_id}")
    print(f"Start Index: {file_path_annotation.start_index}")
    print(f"End Index: {file_path_annotation.end_index}")

Here is an example to use get_file_content:

from pathlib import Path

async def save_file_content(client, file_id: str, file_name: str, target_dir: Optional[Union[str, Path]] = None):
    # Determine the target directory
    path = Path(target_dir).expanduser().resolve() if target_dir else Path.cwd()
    path.mkdir(parents=True, exist_ok=True)

    # Retrieve the file content
    file_content_stream = await client.get_file_content(file_id)
    if not file_content_stream:
        raise RuntimeError(f"No content retrievable for file ID '{file_id}'.")

    # Collect all chunks asynchronously
    chunks = []
    async for chunk in file_content_stream:
        if isinstance(chunk, (bytes, bytearray)):
            chunks.append(chunk)
        else:
            raise TypeError(f"Expected bytes or bytearray, got {type(chunk).__name__}")

    target_file_path = path / file_name

    # Write the collected content to the file synchronously
    with open(target_file_path, "wb") as file:
        for chunk in chunks:
            file.write(chunk)

Teardown

To remove resources after completing tasks, use the following functions:

# Delete the file when done
project_client.agents.delete_vector_store(vector_store.id)
print("Deleted vector store")

project_client.agents.delete_file(file_id=file.id)
print("Deleted file")

# Delete the agent when done
project_client.agents.delete_agent(agent.id)
print("Deleted agent")

Evaluation

Evaluation in Azure AI Project client library is designed to assess the performance of generative AI applications in the cloud. The output of Generative AI application is quantitively measured with mathematical based metrics, AI-assisted quality and safety metrics. Metrics are defined as evaluators. Built-in or custom evaluators can provide comprehensive insights into the application's capabilities and limitations.

Evaluator

Evaluators are custom or prebuilt classes or functions that are designed to measure the quality of the outputs from language models or generative AI applications.

Evaluators are made available via azure-ai-evaluation SDK for local experience and also in Evaluator Library in Azure AI Foundry for using them in the cloud.

More details on built-in and custom evaluators can be found here.

Run Evaluation in the cloud

To run evaluation in the cloud the following are needed:

  • Evaluators
  • Data to be evaluated
  • [Optional] Azure Open AI model.
Evaluators

For running evaluator in the cloud, evaluator ID is needed. To get it via code you use azure-ai-evaluation

# pip install azure-ai-evaluation

from azure.ai.evaluation import RelevanceEvaluator

evaluator_id = RelevanceEvaluator.id
Data to be evaluated

Evaluation in the cloud supports data in form of jsonl file. Data can be uploaded via the helper method upload_file on the project client.

# Upload data for evaluation and get dataset id
data_id, _ = project_client.upload_file("<data_file.jsonl>")
[Optional] Azure OpenAI Model

Azure AI Foundry project comes with a default Azure Open AI endpoint which can be easily accessed using following code. This gives you the endpoint details for you Azure OpenAI endpoint. Some of the evaluators need model that supports chat completion.

default_connection = project_client.connections.get_default(connection_type=ConnectionType.AZURE_OPEN_AI)
Example Remote Evaluation
import os
from azure.ai.projects import AIProjectClient
from azure.identity import DefaultAzureCredential
from azure.ai.projects.models import Evaluation, Dataset, EvaluatorConfiguration, ConnectionType
from azure.ai.evaluation import F1ScoreEvaluator, RelevanceEvaluator, HateUnfairnessEvaluator


# Create project client
project_client = AIProjectClient.from_connection_string(
    credential=DefaultAzureCredential(),
    conn_str=os.environ["PROJECT_CONNECTION_STRING"],
)

# Upload data for evaluation and get dataset id
data_id, _ = project_client.upload_file("<data_file.jsonl>")

deployment_name = "<deployment_name>"
api_version = "<api_version>"

# Create an evaluation
evaluation = Evaluation(
    display_name="Remote Evaluation",
    description="Evaluation of dataset",
    data=Dataset(id=data_id),
    evaluators={
        "f1_score": EvaluatorConfiguration(
            id=F1ScoreEvaluator.id,
        ),
        "relevance": EvaluatorConfiguration(
            id=RelevanceEvaluator.id,
            init_params={
                "model_config": default_connection.to_evaluator_model_config(
                    deployment_name=deployment_name, api_version=api_version
                )
            },
        ),
        "violence": EvaluatorConfiguration(
            id=ViolenceEvaluator.id,
            init_params={"azure_ai_project": project_client.scope},
        ),
    },
)


evaluation_response = project_client.evaluations.create(
    evaluation=evaluation,
)

# Get evaluation
get_evaluation_response = project_client.evaluations.get(evaluation_response.id)

print("----------------------------------------------------------------")
print("Created evaluation, evaluation ID: ", get_evaluation_response.id)
print("Evaluation status: ", get_evaluation_response.status)
if isinstance(get_evaluation_response.properties, dict):
    print("AI Foundry URI: ", get_evaluation_response.properties["AiStudioEvaluationUri"])
print("----------------------------------------------------------------")

NOTE: For running evaluators locally refer to Evaluate with the Azure AI Evaluation SDK.

Tracing

You can add an Application Insights Azure resource to your Azure AI Foundry project. See the Tracing tab in your AI Foundry project. If one was enabled, you can get the Application Insights connection string, configure your Agents, and observe the full execution path through Azure Monitor. Typically, you might want to start tracing before you create an Agent.

Installation

Make sure to install OpenTelemetry and the Azure SDK tracing plugin via

pip install opentelemetry
pip install azure-ai-projects azure-identity opentelemetry-sdk azure-core-tracing-opentelemetry

You will also need an exporter to send telemetry to your observability backend. You can print traces to the console or use a local viewer such as Aspire Dashboard.

To connect to Aspire Dashboard or another OpenTelemetry compatible backend, install OTLP exporter:

pip install opentelemetry-exporter-otlp

How to enable tracing

Here is a code sample that shows how to enable Azure Monitor tracing:

from opentelemetry import trace
from azure.monitor.opentelemetry import configure_azure_monitor

# Enable Azure Monitor tracing
application_insights_connection_string = project_client.telemetry.get_connection_string()
if not application_insights_connection_string:
    print("Application Insights was not enabled for this project.")
    print("Enable it via the 'Tracing' tab in your AI Foundry project page.")
    exit()
configure_azure_monitor(connection_string=application_insights_connection_string)

scenario = os.path.basename(__file__)
tracer = trace.get_tracer(__name__)

with tracer.start_as_current_span(scenario):
    with project_client:

In addition, you might find helpful to see the tracing logs in console. You can achieve by the following code:

project_client.telemetry.enable(destination=sys.stdout)

How to trace your own functions

The decorator trace_function is provided for tracing your own function calls using OpenTelemetry. By default the function name is used as the name for the span. Alternatively you can provide the name for the span as a parameter to the decorator.

This decorator handles various data types for function parameters and return values, and records them as attributes in the trace span. The supported data types include:

  • Basic data types: str, int, float, bool
  • Collections: list, dict, tuple, set
    • Special handling for collections:
      • If a collection (list, dict, tuple, set) contains nested collections, the entire collection is converted to a string before being recorded as an attribute.
      • Sets and dictionaries are always converted to strings to ensure compatibility with span attributes.

Object types are omitted, and the corresponding parameter is not traced.

The parameters are recorded in attributes code.function.parameter.<parameter_name> and the return value is recorder in attribute code.function.return.value

Troubleshooting

Exceptions

Client methods that make service calls raise an HttpResponseError exception for a non-success HTTP status code response from the service. The exception's status_code will hold the HTTP response status code (with reason showing the friendly name). The exception's error.message contains a detailed message that may be helpful in diagnosing the issue:

from azure.core.exceptions import HttpResponseError

...

try:
    result = project_client.connections.list()
except HttpResponseError as e:
    print(f"Status code: {e.status_code} ({e.reason})")
    print(e.message)

For example, when you provide wrong credentials:

Status code: 401 (Unauthorized)
Operation returned an invalid status 'Unauthorized'

Logging

The client uses the standard Python logging library. The SDK logs HTTP request and response details, which may be useful in troubleshooting. To log to stdout, add the following:

import sys
import logging

# Acquire the logger for this client library. Use 'azure' to affect both
# 'azure.core` and `azure.ai.inference' libraries.
logger = logging.getLogger("azure")

# Set the desired logging level. logging.INFO or logging.DEBUG are good options.
logger.setLevel(logging.DEBUG)

# Direct logging output to stdout:
handler = logging.StreamHandler(stream=sys.stdout)
# Or direct logging output to a file:
# handler = logging.FileHandler(filename="sample.log")
logger.addHandler(handler)

# Optional: change the default logging format. Here we add a timestamp.
#formatter = logging.Formatter("%(asctime)s:%(levelname)s:%(name)s:%(message)s")
#handler.setFormatter(formatter)

By default logs redact the values of URL query strings, the values of some HTTP request and response headers (including Authorization which holds the key or token), and the request and response payloads. To create logs without redaction, add logging_enable = True to the client constructor:

project_client = AIProjectClient.from_connection_string(
    credential=DefaultAzureCredential(),
    conn_str=os.environ["PROJECT_CONNECTION_STRING"],
    logging_enable = True
)

Note that the log level must be set to logging.DEBUG (see above code). Logs will be redacted with any other log level.

Be sure to protect non redacted logs to avoid compromising security.

For more information, see Configure logging in the Azure libraries for Python

Reporting issues

To report an issue with the client library, or request additional features, please open a GitHub issue here. Mention the package name "azure-ai-projects" in the title or content.

Next steps

Have a look at the Samples folder, containing fully runnable Python code for synchronous and asynchronous clients.

Explore the AI Starter Template. This template creates an Azure AI Foundry hub, project and connected resources including Azure OpenAI Service, AI Search and more. It also deploys a simple chat application to Azure Container Apps.

Contributing

This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.microsoft.com.

When you submit a pull request, a CLA-bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., label, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information, see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.