Enable human interaction in AutoGenStudio - Solution 2 #3476
Conversation
Update *WorkflowManager* classes: - Add async `a_send_message_function` parameter to mirror `send_message_function` param. - Add async `a_process_message` coroutine to mirror the synchronous `process_message` function. - Add async `a_run` coroutine to mirror the `run` function - Add async `_a_run_workflow` coroutine to mirror the synchronous `_run_workflow` function. Update *ExtendedConversableAgent* and *ExtendedGroupChatManager* classes: - Override the async `a_receive` coroutines Update *AutoGenChatManager*: - Add async `a_send` and `a_chat` coroutines to mirror their sync counterparts. - Accept the `WebSocketManager` instance as a parameter, so it can do Async comms directly. Update *app.py* - Provide the `WebSocketManager` instance to the *AutoGenChatManager* constructor - Await the manager's `a_chat` coroutine, rather than calling the synchronous `chat` function.
Updates to *ExtendedConversableAgent* and *ExtendedGroupChatManager* classes - override the `get_human_input` function and async `a_get_human_input` coroutine Updates to *WorkflowManager* classes: - add parameters `a_human_input_function` and `a_human_input_timeout` and pass along on to the ExtendedConversableAgent and ExtendedGroupChatManager - fix for invalid configuration passed from UI when human input mode is not NEVER and no model is attached Updates to *AutoGenChatManager* class: - add parameter `human_input_timeout` and pass it along to *WorkflowManager* classes - add async `a_prompt_for_input` coroutine that relies on `websocket_manager.get_input` coroutine (which snuck into last commit) Updates to *App.py* - global var HUMAN_INPUT_TIMEOUT_SECONDS = 180, we can replace this with a configurable value in the future
|
@victordibia the main reason the other solution didn't work the same in different machines is because it relied on running two separate asyncio event loops in two separate threads. Some googling around suggested doing so is unsupported, and leads to undefined behaviors (due to interaction with the GIL). This method should work for everyone because it uses only the main asyncio event loop for getting user input. To make that work, I had to convert the call to AutoGen's |
There was a problem hiding this comment.
Just tested this @SailorJoe6 , works as advertised! I learned a thing or two.
This PR also provides building blocks for many other good ideas e.g., using the same socket style messaging to interrupt or stop agents.
samples/apps/autogen-studio/frontend/src/components/views/playground/chatbox.tsx
Show resolved
Hide resolved
Why are these changes needed?
The AutoGenStudio UI doesn't currently support user input modes 'ALWAYS', or 'TERMINATE' which are vital when allowing agents to generate and run code, so that the user can verify the code before it gets execute in their environment. This is likely due to difficulties encountered with Python's asyncio event loop getting blocked by synchronous calls to the AutoGen chat functions. This PR fixes those issues and adds support for all human input modes:
This code comes in multiple commits so you can follow the logical progression through:
Related issue number
Resolves #1358
Resolves #1664
Part of roadmap #737
Checks