Part 1: AutoGen Unleashed: Crafting Intelligent Group Chats from Scratch

Introduction

AutoGen simplifies the creation and management of autonomous agents for various tasks. This article demonstrates setting up a multi-agent system to fetch real-time stock prices, showcasing AutoGen's flexibility and power.

The image shows AutoGen's features: individual conversable agents and their ability to engage in multi-agent chats, either jointly or hierarchically, allowing for complex conversations and agent customization. 

Setting Up the Environment

A virtual environment isolates project dependencies, ensuring a clean workspace. We'll use Python's venv to create one and then install AutoGen along with other necessary packages.

python -m venv venv
source venv/bin/activate
pip install pyautogen

Initial Configuration and File Setup

Now, onto the code. Create a file named app.py. This file will serve as the heart of our AutoGen setup. We begin by importing the necessary libraries from AutoGen, which will facilitate the creation of our chat agents and manage our configurations.

from autogen import AssistantAgent, UserProxyAgent, GroupChat, GroupChatManager, config_list_from_json

Adding Search Functionality

A crucial feature of our setup is the ability to perform web searches. This is achieved through a custom search function, declared and implemented to fetch data as needed. Our setup includes web searches via a custom function using serper.dev, offering quick data retrieval and streamlined integration. Let's create a new file search.py:

# search.py

from datetime import datetime, timedelta
import json
import requests
import os

search_declaration = {
   "name": "Search",
   "description": "Search at Google and returns Title, Link, and Snippet",
   "parameters": {
     "type": "object",
     "properties": {
     "query": {
       "type": "string",
       "description": "The search query"
     }
   },
 "required": ["query"]
 },
}

def search(query):
  # Calculate the date 6 months ago from today
  six_months_ago = datetime.now() - timedelta(days=36*30) # approximating 3 years
  date_str = six_months_ago.strftime('%Y-%m-%d')

  # Append the date filter to the query
  query = f"{query} after:{date_str}"

  url = "https://google.serper.dev/search"

  # Get the API key from an environment variable
  api_key = os.environ.get('GOOGLE_SERPER_API_KEY')

  if not api_key:
   raise ValueError(
     "Environment variable 'GOOGLE_SERPER_API_KEY' not set!"
  )

  payload = json.dumps({
    "q": query
  })

  headers = {
    'X-API-KEY': api_key,
    'Content-Type': 'application/json'
  }

  response = requests.request("POST", url, headers=headers, data=payload)

  return response.json()

Integrating this function into our Assistant Agent allows it to utilize external data sources, enhancing its utility. For this we first have to import the new function into our app.py:

from search import search, search_declaration

Implementing the Assistant Agent

Our first agent, the Assistant Agent, acts as the primary responder within our chat setup.
Here's how to initialize it:

assistant = AssistantAgent('assistant', llm_config={"functions": [search_declaration], "config_list": config_list_from_json("OAI_CONFIG_LIST")})

This snippet assigns a name to our agent and configures it with a language model, ready to handle tasks and queries.

The file OAI_CONFIG_LIST will define the models and API keys used by our AutoGen setup:

[{
 "model": "gpt-4",
 "api_key": "YOUR OPENAI API KEY"
}]

Setting Up the User Proxy Agent

The User Proxy Agent stands between the user and the Assistant, mediating interactions. It's configured to execute code and facilitate user-agent dialogues. 



user_proxy = UserProxyAgent('user_proxy', function_map={"Search": search}) 

This configuration allows the User Proxy Agent to access custom functions, such as our search functionality.

llm_config = {"config_list": config_list_from_json("OAI_CONFIG_LIST")}
cynic = AssistantAgent('cynic', system_message="You are a cynic and you comment everything", llm_config=llm_config) 

groupchat = GroupChat(agents=[assistant, cynic, user_proxy], messages=[])

groupchat_manager = GroupChatManager(groupchat, llm_config=llm_config) 

user_proxy.initiate_chat(groupchat_manager, message="Find the latest stock price of Apple")

export AUTOGEN_USE_DOCKER=false
python app.py