Find a live demo here! A chat assistant to help users design experimental drug discovery workflows using Ersilia models with ease!

Below is a schema of how Ersilia Assistant development is guided

This project requires Python version 3.10 or higher.

To get started, clone this repository and create a virtual environment, and install the dependencies

Note: Install the package in editable mode so you can easily review changes

source venv/bin/activate # Or your environment manager of choice

git clone https://github.com/ersilia-os/ersilia-assistant.git
cd ersilia-assistant
pip install -e .

Make sure you have installed a llamafile of your choice, for example, Llama 3.1 8B-Instruct. In our experience, Meta-Llama-3.1-8B-Instruct.Q5_K_M performs better among the 8B versions available on HuggingFace.

TL;DR on quantization: Q5 quantization is better than Q4, but slower, similarly, M is better than S, and slower.

In a separate shell, run the following:

chmod +x /path/to/llamafile # You need to make the llamafile executable only once
./path/to/llamafile --server --nobrowser

Note: This is a blocking operation, and you should not close this shell. Optionally, you can use nohup to run this as a background process and exit the shell:

nohup ./path/to/llamafile --server --nobrowser &

You can verify that the llamafile server is running by opening a browser window and navigating to http://localhost:8080/

To deploy this entire setup locally, we have a simple script that creates necessary directories and installs the required assets as needed. To do this, run the following:

cd bin
./local-build

Since this project requires Python version 3.10 or higher, make sure that is the default Python version running in your shell.