A continuation of the LLM hackathon project, yeLLowhaMMer: A Multi-modal Tool-calling Agent for Accelerated Research Data Management.

This repository will explore using the Jupyter AI plugin to provide an agentic interface to datalab, with the idea of having this as an additional UI that datalab users can use to interact with their data, either deployed for their instance, or run locally.

Initial development tasks

• [ ] Reproduce the hackathon project with the Jupyter AI plugin, probably by extending the %%ai cell magic to something like %%yellowhammer that includes the yellowhammer system prompt, and guides through the registration of a datalab API key and any provided LLM API keys.
• [ ] Use yellowhammer to generate API examples for the underlying datalab-python-api package.
• [ ] Consider deploying this as a JupyterHub that datalab instances can link to directly.
• [ ] Integrate the results much more closely into datalab itself, i.e., attaching notebooks to the relevant samples, and recording the provenance of AI generated data recording.

Installation

This repository uses uv for the entire packaging workflow. Once you have installed uv following their documentation, you can install this repository by cloning and running uv sync in the root directory (optionally with --dev if you plan to develop it further).

git clone git@github.com:datalab-org/yellowhammer
cd yellowhammer
uv sync --dev

Launching example notebooks

You can launch the example notebook locally with uv too:

uv run jupyter lab examples/

The examples will require you to bring your own datalab API key (for your instance of choice) and API keys for any underlying LLM providers (OpenAI, Anthropic, etc.).

These can be set in your shell profile, or simply in your shell before launching Jupyter, using:

export OPENAI_API_KEY=sk-proj...
export ANTHROPIC_API_KEY=sk-ant...

yellowhammer by default will come preloaded with the relevant OpenAI and Anthropic packages. You can see how to configure other providers in the Jupyter AI plugin documentation.