This is an example of chatgpt-like bot. It is simplified, but demonstrates the core concepts nicely.
The core idea is that we:
- Accept a prompt
- Check to see if the prompt is "safe" (this is currently a no-op, but its there for show)
- Decide the response mode by querying openAI (answer a question, generate an image, write some code, and can't decide/prompt for more)
- Query one of our models with a modified prompt to return the right response
- Return the response
- Go back to the beginning and wipe the state.
We have both a hamilton and non-hamilton implementation of this for comparison. For the simplicity of this example, Hamilton is likely overkill. There are a few nodes that could leverage it well, but the rest are too simple to worry about putting in a DAG. As examples get more complex, we anticipate that Hamilton nodes will play a more important role.
The state machine this encapsulates takes the following form:
Note that it is currently wired up to handle errors in the state machine, but the actions do not respect that yet. Coming soon.
We have three files:
- application.py -- This contains a mainline to generate the graph portrayal.
- requirements.txt -- Just the requirements. All this needs is Burr/Streamlit/openai
- streamlit_app.py -- This contains a simple Streamlit app to interact with the cow
- application_with_traces.py -- This showcases how to use the tracing feature.
You must have your
OPENAI_API_KEYset in your environment. You can do this by running.
To run on streamlit, you can launch the app:
streamlit run streamlit_app.pyThis has the state machihne displayed as well as a quick debugging/demo. You should be using the UI, however, as it is more powerful and responsive. You can run the simple app with:
streamlit run simple_streamlit_app.py & burrThis will open up both a streamlit app and a server at the same time.