autobots: AutoGen agent workflow design

🔁 Steps to Run

1. It is recommended to create a virtual environment using python -m venv {newenv}
2. Run pip install -r requirements.txt
3. Set export PYTHONPATH=$(pwd) (and if needed, export PATH={which_python_path})
4. Set the OPENAI_API_KEY environment variable using export OPENAI_API_KEY="your_api_key", then reload using source ~/.bashrc
5. Change configurations in the src/web/config.json file as needed
6. Ensure that Docker is installed
7. In one terminal, run python src/__main__.py
8. In another terminal, make post requests with test Python scripts in src/web/testing/, for example: python src/web/testing/{test_name}.py

🕵️ AutoGen Agents and Chat Management

autogenchat.py

•   class AutoGenChatManager

  • Handles the automated generation and management of chat interactions.
  • An AutoGenChatManager instance is initialized with a message queue that organizes the messages to be sent between assistant agents and the user proxy agent.
  • Initializes workflow managers based on incoming workflow configuration.
  • Outputs a summary of the conversation, including the user's and assistant agents' messages, token cost and usage, as well as the start and end time.

datamodel.py

Provides the foundational data structures to be used by workflow managers and the chat manager.

• Dataclasses: used for storing data, such as the Message, Session, Gallery, and AgentWorkFlowConfig classes.
• Pydantic Model Structures: handle data in APIs, applications, and configurations (e.g. Agent, Workflow).
• Enum classes: define categories or options a variable can take, like AgentType, CodeExecutionConfigTypes (locally or via Docker container), and WorkFlowType.
• Additional Functionality: __post_init__ for additional initialization after the automatic __init__ method and classes to convert between class instances, JSON, and dictionaries.

workflowmanager.py

Autonomous vs. Sequential Workflow

• autonomous: agents operate independently and simultaneously without waiting for other agents to complete their tasks.
• sequential: agents operate in a specific pre-determined order, where each agent waits for the previous one to complete its task.

Assistant vs. User Proxy

• assistant: acts as a helper or advisor to answer queries and perform tasks input by the user.
• userproxy: acts as a proxy for a human user, responsible for asking questions.

Group Chat Setting

In the groupchat setting, multiple agents can commmunicate with each other aside from the user. The order in which agents answer the query is self-determined.

🗄 SQLite Database Operations

Autobots supports user message and session addition, retrieval, and deletion (including database clearing). When app.py (the file with API GET and POST endpoints) is run, a database.sqlite SQLite database is created with 3 tables: messages, sessions, and gallery. The messages table stores one entry for each user query and one entry for each combined assistant(s) response, and each message item is uniquely identified by its msg_id. Each message item belongs to a session uniquely identified in the sessions table by the session_id. A session can be published to the gallery table where each gallery item is uniquely identified by its gallery_id.

Database functions (more details in src/utils/utils.py and src/web/app.py)

•   load_messages(user_id: str, session_id: str, dbmanager: DBManager) -> List[Message]:

  • Load messages from the database for a given user and session specified by their IDs.
  • Used when retrieving user history or publishing a session to the gallery database.

•   save_message(message: Message, dbmanager: DBManager) -> None:

  • Save a message to the database.
  • Used when adding a message to the database.

•   delete_message(user_id: str, msg_id: Optional[str], session_id: Optional[str], dbmanager: DBManager, delete_all: bool) -> []:

  • Delete a message from the database. delete_all is set to True when clearing all messages belonging to a session.
  • Used when deleting an individual message or clearing the database.

•   get_sessions(user_id: str, dbmanager: DBManager) -> List[Session]:

  • Retrieve all sessions for a given user from the database.
  • Used when creating a session, retrieving a session, or clearing the database.

•   create_session(user_id: str, session: Session, dbmanager: DBManager) -> List[Session]:

  • Create a new user session in the database.

•   delete_user_sessions(user_id: str, session_id: str, dbmanager: DBManager) -> List[Session]:

  • Delete a specific session from the database.
  • Used when clearing the database.

•   publish_session(session: Session, tags: Optional[List[str]], dbmanager: DBManager) -> Gallery:

  • Publish a user's session to the gallery.

•   get_gallery(gallery_id: Optional[str], dbmanager: DBManager) -> List[Gallery]:

  • Retrieve gallery items from the database.

🧪 Testing

See the src/web/testing/ folder to run API endpoint tests. Workflow configurations are outlined in the test files.

🚶‍♂️ Next Steps

One area to explore is adding skills to agents in order to enable external tool usage. More database functions, such as more gallery operations, can be added to existing utilities.

💡 References

Many thanks to AutoGen for providing the framework.