Closed piopy closed 8 months ago
The instruction following ability for gpt-3.5-turbo is very weak. The generated name cannot be guaranteed to follow an expected format when using gpt-3.5 series model. To ensure the quality of agents, I highly recommend using gpt-4 (or gpt-4-1106-preview) instead as the build_manager. For this issue, I will add a trigger to replace the space as underline later.
Describe the bug
AgentBuilder creates agents that contains spaces in their names.
When initializing the conversation, I receive the following error:
BadRequestError: Error code: 400 - {'error': {'message': "' Content_Generator' does not match '^[a-zA-Z0-9_-]{1,64}$' - 'messages.3.name'", 'type': 'invalid_request_error', 'param': None, 'code': None}}
Steps to reproduce
Step 1: install dependencies Step 2: write API key and model (gpt-3.5-turbo) in a json for the autobuilder. Set also the llm_config dictionary for the conversation. (same gpt used) Step 3: Set AutoBuilder and write the building task Step 4: Create agent_list, agent_configs using builder.build(building_task, llm_config) Step 5: Create a GroupChat and a manager (same llm_config) Step 6: Initiate conversation with agent_list[0] to the manager defining a problem. Error will come if AutoBuilder generated wrong formatted names.
Expected Behavior
Autobuilt agents conversating each other
Screenshots and logs
Additional Information
openai == 1.3.8 pyautogen == 0.2.2 (pip install git+https to this repo)
I used a workaround in the AutoBuilder's prompt, adding "Generate agents without spaces in their names." command