Use popular Python AI agent frameworks with GitHub Models or Azure OpenAI models!

[pamelafox]

Introduction

This repo includes multiple examples of single agents and multi-agents coded with popular Python AI agent frameworks:
https://github.com/Azure-Samples/python-ai-agent-frameworks-demos

Frameworks included: openai-agents, Semantic Kernel, AutoGen, Langgraph, Pydantic AI, smolagents, llamaindex

You can either open the repo in GitHub Codespaces and immediately use the models with GitHub Models, or you can use "azd up" to create an Azure OpenAI model, and run the examples with that model.

The examples are also translated into spanish in the spanish/ folder.

We put on two live streams highlighting the examples:

Prototyping AI Agents with GitHub Models (English)

• Recording on YouTube
• Slides

Prototipando agentes de IA con GitHub Models (Español)

Grabación en YouTube
Diapositivas (PPT)

Features and Screenshots

These are some of the architectures demonstrated in the examples:

Technical Details

Technologies used: Python, openai-agents, Semantic Kernel, AutoGen, Langgraph, Pydantic AI, smolagents, llamaindex, rich

Architecture overview: GitHub Models or Azure OpenAI

Challenges and Solutions

It was tricky to understand how to implement the same scenario across multiple frameworks, but it was also fun to figure that out - helped me understand the frameworks better.

[amynic]

Wow! thanks for sharing @pamelafox this is super useful for those getting started with Agent frameworks and Azure AI Foundry. I need to take a look at smolagents will have a go!

[nicholasdbrady]

+1 as a Python dev, I love how smolagents works natively with Gradio

..I think I need to learn Pydantic AI next, thanks @pamelafox!

[chanezon]

Thank you @pamelafox, great to see how the different frameworks stack up for the same scenario!

[Kamal-Moha]

Hi @pamelafox,

I have followed your code & using pydanticAI with Github models. I have noticed Github models are actually NOT making function/tool calls, but only using system_prompts which is not good for my purpose.
I have compared the Github Models with a Mistral Models(not via Github). I have clearly seen that the Mistral model is making the tool calls and providing better replies/answers to the end user.

Here is me comparing the Mistral model vs Github model in making tool calls for my purpose

Mistral Model

from pydantic_ai import Agent
from pydantic_ai.models.mistral import MistralModel
# from pydantic_ai.providers.mistral import MistralProvider
import logfire

logfire.configure()
logfire.instrument_asyncpg()

mistral_model = MistralModel('mistral-large-latest')

Github Model

import asyncio
import os

# import azure.identity
from dotenv import load_dotenv
from openai import AsyncAzureOpenAI, AsyncOpenAI
from pydantic_ai import Agent
from pydantic_ai.models.openai import OpenAIModel
from pydantic_ai.providers.openai import OpenAIProvider

# Setup the OpenAI client to use either Azure OpenAI or GitHub Models
load_dotenv(override=True)

client = AsyncOpenAI(api_key=os.environ["GITHUB_TOKEN"], base_url="https://models.inference.ai.azure.com")
github_model = OpenAIModel('MAI-DS-R1', provider=OpenAIProvider(openai_client=client))

Here's my agent using either the Mistral Model or the Github Model

from __future__ import annotations as _annotations

import asyncio
import os
from dataclasses import dataclass
from typing import Any

import logfire
from devtools import debug
from httpx import AsyncClient

from pydantic_ai import Agent, ModelRetry, RunContext

logfire.configure()
logfire.instrument_asyncpg()


@dataclass
class Deps:
    client: AsyncClient
    auth_key: str | None


async def get_trans_history(
    ctx: RunContext[Deps], phone_number: str
) -> dict[str, str]:
    """Get all the transactions done by a number(user).

    Args:
        ctx: The context.
        phone_number: Phone Number for the user.
    """
    if ctx.deps.auth_key is None:
        # if no API key is provided, return a dummy response (London)
        return "No API Key has been provided. I need an API Key to do a search"


    headers = {
        'accept': '*/*',
        'accept-language': 'en-US,en;q=0.9',
        'authorization': ctx.deps.auth_key,
        'content-type': 'application/json'
    }

    params = {
        'limit': '100',
        'offset': '0',
        'show_all': 'true',
    }
    with logfire.span('calling SEL Support API', params=params) as span:
        r = await ctx.deps.client.get(f'https://witaxiapi.com/wallet/{phone_number}/transactions',
                                      headers=headers,
                                      params=params
                                      )
        r.raise_for_status()
        data = r.json()
        span.set_attribute('response', data)

    if data:
        print(data)
        return data
    else:
        raise ModelRetry(f'Could not find the transaction for the phone number: {phone_number}')


witaxipay_agent = Agent(
    # mistral_model,  # picking the mistral model
    github_model, # picking the github model
    tools=[get_trans_history],
    system_prompt=(
        'You are an assistant for a payment(wallet) app that helps riders pay for their fares easily'
        'Be knowledgeable and smart about technical issues that customers have'
        'Have willingness to help and be as detailed as possible'
        'Use the `get_trans_history` tool to get the transaction history done by the number. This shows every transaction done by the user. But understand the specific issues that the user is asking for'
        'Find the date(s) & that the user has been having this issue'
    ),
    deps_type=Deps,
    retries=2,
    instrument=True,
)

# Calling the Agent
async def main():
    async with AsyncClient() as client:
      deps = Deps(
          client=client,
          auth_key= 'bearer 48f6g78-d2cc-41a1-965b-c****f'
      )
      res = await witaxipay_agent.run("I'm a user with phone number 835163582. My wallet topups have failed, I don't know why. Please help", deps=deps)
      
      debug(res)
      print('Response:', res.output)

if __name__ == '__main__':
    asyncio.run(main())

When I use mistral_model in the Agent, I see in the logfire logs that my tool get_trans_history is being called. Here are few images from logfire and my terminal showing the get_trans_history being called

This leads to a good response from the Agent when using the mistral model.

But when I use the github_model, I see in the in the logfire logs that the agent is NOT calling my tool get_trans_history. Here are a few images to show that;

As you can see, the agent(when using the github_model) is going from; user interaction directly to Assistant response without any tool call as you can see in the images.

Please help me as I'm not able to make tool calls in PydanticAI when using any Github Models. Tell me where I'm going wrong. Thanks

[pamelafox]

I have added an example to the agents frameworks repo to verify that Pydantic AI can use tool calling with GitHub Models:
https://github.com/Azure-Samples/python-ai-agent-frameworks-demos/blob/main/examples/pydanticai_tools.py

I used gpt-4o there, since it supports function calling. In your example, you're using "MAI-DS-R1", a model which has not been trained for function calling. You can see what models support function calling on this page:
https://learn.microsoft.com/azure/ai-foundry/model-inference/concepts/models#microsoft
Notice it says "Tool calling: No" under that model.

[Kamal-Moha]

Thanks a lot @pamelafox for the clarification.

I will ensure to test with Github models that have function calling capability.