Vanilla AI Agents

Lightweight library demonstrating how to create agenting application without using any specific framework.

Table of Contents

• Features
• Future work
• Getting Started
• Demos
• Testing
• License
• Contributing

Features

This project framework provides the following features:

• Multi-agent chat with several orchestration options
  • Dynamic routing (including option to look for available tools to decide)
  • Beforehand planning with optional repetition via feedback loop
• Agent state management
• Custom stop conditions
• Interactive or unattended user input
• Chat resumability
• Function calling on agents
• Constrained agent routing
• Sub-workflows
• Simple RAG via function calls
• Image input support
• Ability to run pre and post steps via Sequence
• Conversation context "hidden" variables, which are not displayed to the user but agents can read and write to access additional information
• Usage metrics tracking per conversation, plus internal log for debuggability
• Multiple strategies for agent to filter conversation messages (All, last N, top K and Last N, summarize, etc..)
• LLMLingua (extras module) support to compress system prompts via strategies
• LLM support for Structured Output
• DAPR integration
  • Native submodule to host Askable and Workflow as Dapr Actors
  • Dapr PubSub integration for Workflow to enable
    • Event sourcing
    • Decoupled communication between workflows
• Multi-agent chat with multiple users
  • Dapr PubSub integrations allows to move from one-to-many to many-to-many conversations, with different User instances impersonating different user profiles
  • Demo Repo (Coming soon)
• Remoting support ((remote module)), allowing agents to be run on a remote server and accessed elsewhere
  • REST and gRPC channels supported
  • Default implementation to run hosts with agent discovery and registration
• Generated Code execution locally and via ACA Dynamic Sessions
• Streaming support, even over REST or gRPC agents

Future work

• Plugins
  • Azure AI Search plugin
  • DB plugin
  • API plugin

Getting Started

Prerequisites

Python 3.11 or later is required to run this project.

Quickstart

git clone https://github.com/Azure-Samples/vanilla-aiagents

cd "vanilla-aiagents"

# Create a virtual environment
python -m venv .venv

# Activate the virtual environment

# On Windows
.\.venv\Scripts\activate
# On Unix or MacOS
source .venv/bin/activate

# Install the required dependencies
pip install -r requirements.txt

# Clone .env.sample to .env and update the values
cp .env.sample .env

Here is a simple example of how to use the framework:

import os
from vanilla_aiagents.llm import AzureOpenAILLM
from vanilla_aiagents.agent import Agent
from vanilla_aiagents.team import Team
from vanilla_aiagents.workflow import Workflow

llm = AzureOpenAILLM({
    "azure_deployment": os.getenv("AZURE_OPENAI_MODEL"),
    "azure_endpoint": os.getenv("AZURE_OPENAI_ENDPOINT"),
    "api_key": os.getenv("AZURE_OPENAI_KEY"),
    "api_version": os.getenv("AZURE_OPENAI_API_VERSION"),
})

# Initialize agents and team
sales = Agent(id="sales", llm=llm, description="A sales agent", system_message="""
You are a sales assistant. You provide information about our products and services.

# PRODUCTS
- Product 1: $100, description
- Product 2: $200, description
- Product 3: $300, description
""")
support = Agent(id="support", llm=llm, description="A support agent", system_message="""
You are a support assistant. You provide help with technical issues and account management.

# SUPPORT GUIDELINES
- For technical issues, please provide the following information: ...
- For account management, please provide the following information: ...
""")
team = Team(id="team", description="Contoso team", members=[sales, support], llm=llm)

# Create a workflow
workflow = Workflow(askable=team)

# Run the workflow
result = workflow.run("Hello, I'd like to know more about your products.")
print(workflow.conversation.messages)

Submodules

remote

This module provides a way to run agents on a remote server and access them elsewhere. It includes support for REST and gRPC channels, as well as a default implementation to run hosts with agent discovery and registration.

Additionally, it features Dapr integration, allowing for the hosting of Askable and Workflow as Dapr Actors, enabling event sourcing and decoupled communication between workflows.

See the Remote Agents documentation and Actors documentation for more information.

extras

This module provides additional features to enhance the functionality of the framework. It includes support for:

• LLMLingua to compress system prompts

Demos

notebooks folder contains a few demo notebooks that demonstrate how to use the framework in various scenarios.

Testing

To run the tests, execute the following command:

invoke test

To run run selected tests, execute the following command:

invoke test --test-case <test_case>

Testing also includes code coverage.

Building

To build the project, run the following command:

invoke build --version <version>

Output wheel will be available in the dist folder under vanilla_aiagents with naming vanilla_aiagents-<version>-py3-none-any.whl.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Contributing

We welcome contributions! Please see CONTRIBUTING.md for details on how to contribute.