Visualizing AutoGen workflow

[marklysze]

-- Draft PR --

Creates a visual diagram of an AutoGen workflow! Check the flow between agents, LLM calls, function calls.

This is a starting point and your feedback, contributions, and testing are needed! Recreating a workflow from logs isn't always straightforward... :)

The proposed technical approach is:

• Use existing AutoGen logging mechanism as the source to build the diagram
• Utilise GraphViz's diagram building capability. More info on GraphViz.
• Create an output to SVG (or a more interactive format that allows you to view underlying data)

Steps:

• Add additional logging points to help give necessary detail (such as when in a nested chat, when in group chat select speaker, max turns reached, etc.)
• Add new tool/package/contrib that creates the visualisation. It will be a sub-package referenced as autogen.visualize and will be installed as an optional dependency pip install autogen[visualize].

Work to do:

• Add new logging events (this PR, done some)
• Add support for file logging (this PR, done)
• Add new tool/package/contrib for the actual visualisation logic (this PR, done some)
• Tests (this PR, not started)
• Documentation, examples, installation guide (basic guide in this PR)

Future (not this PR):

• Add support for SQLite logger
• Documentation / Blog

How to run it:

• You will need to install GraphViz twice - install GraphViz for your OS (see here and also install the Python package pip install graphviz. You need both.
• Load this branch

Code examples:

# Create a diagram from an existing log file, saved file as: autogen_logs/test_output.svg

from autogen.visualize import Visualize
visualisation = Visualize("./autogen_logs/20240912_224640_runtime_async_function_call.log")
visualisation.create_svg("autogen_logs/test_output.svg")

# Create a diagram from a file logger object, saved file as: autogen_logs/my_diagram.svg

from autogen import ConversableAgent, UserProxyAgent
import datetime
import autogen.runtime_logging
from autogen.visualize import Visualize

altmodel_llm_config = {
    "config_list":
    [
        {
            "api_type": "ollama",
            "model": "llama3.1:8b-instruct-q8_0",
            "client_host": "http://192.168.0.1:11434",
            "cache_seed": None,
        }
    ]
}

timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
filename = f"{timestamp}_runtime_simple_chat.log"

# Start logging with logger_type and the filename to log to
logging_session_id = autogen.runtime_logging.start(logger_type="file", config={"filename": filename})
print("Logging session ID: " + str(logging_session_id))

# Create the agent that uses the LLM.
assistant = ConversableAgent("agent", llm_config=altmodel_llm_config)

# Create the agent that represents the user in the conversation.
user_proxy = UserProxyAgent("user", code_execution_config=False)

# Let the assistant start the conversation.  It will end when the user types exit.
assistant.initiate_chat(user_proxy, message="How can I help you today?")

autogen.runtime_logging.stop()

visualisation = Visualize(autogen.runtime_logging.autogen_logger) # Uses logger object
visualisation.create_svg("autogen_logs/my_diagram.svg", True)

Note: In VSCode if you preview the SVG diagram you should be able to hover over any bold text and it will show you the message/content.

What does it look like so far

CC for help @Hk669

Why are these changes needed?

Helps developers understand how AutoGen works under the hood. Testing / Production, helps understand running workflows.

Checks

• I've included any doc changes needed for https://autogen-ai.github.io/autogen/. See https://autogen-ai.github.io/autogen/docs/Contribute#documentation to build and test documentation locally.
• I've added tests (if relevant) corresponding to the changes introduced in this PR.
• I've made sure all auto checks have passed.

[Hk669]

@marklysze this looks perfect, let me try to help. thanks for the efforts on making this better.