This project provides a ThoughtCallbackHandler class that extends BaseCallbackHandler for handling callbacks in a Language Model (LM) execution context. It allows you to display the output of an LM along with intermediate results history.
- Clone the repository.
- Install the required dependencies.
To use ThoughtCallbackHandler, simply import the class and create an instance:
from thought_callback_handler import ThoughtCallbackHandler
...
callback_handler = ThoughtCallbackHandler()Then, pass this callback_handler to your LM or agent, and it will handle the callbacks accordingly. For instance:
response = executor.run({"input":payload, "chat_history":chat_history}, callbacks=[callback_handler])The ThoughtCallbackHandler maintains a history of thoughts and codes encountered during the execution of the Language Model. This is accomplished through the history attribute, which is updated each time a new action is processed or when the chain of thought ends. These histories can be helpful for understanding the sequence of operations and decisions made by the Language Model, aiding in debugging and performance optimization.
Though ThoughtCallbackHandler can function standalone, it's also designed with extensibility in mind and can be integrated with interactive tools such as Streamlit.
To leverage Streamlit's real-time updating functionality, you can pass a custom Streamlit callback handler to the ThoughtCallbackHandler. This custom Streamlit callback handler can be designed to take the thought and code histories at each end of the step and display it on a Streamlit component.
This project is licensed under the MIT License.
Contributions are welcome! Please open an issue or submit a pull request for any bugs, improvements, or features.