What is this?
This is a chatbot for health questions that runs entirely on your own computer. You don't need an account, an API key, or an internet connection once it's set up.
It has one important switch, Guardrails:
- Guardrails On. Every message is checked against healthcare safety rules before the AI answers. Unsafe requests are refused, patient identifiers are removed, and risky answers come with warnings.
- Guardrails Off. The message goes straight to the raw AI model with no checks, so you can compare.
Guardrails are safety rules placed around an AI model. They check what goes in and shape what comes out.
Partners
| Partner | Website |
|---|
| WallBridge AI, AI strategy for leadership teams | wallbridge.ai |
| Valix AI, AI security for regulated organizations | valix.ai |
| University of Health and Allied Sciences, Ho | uhas.edu.gh |
Tools used
You don't need to know these tools to run the demo. This table explains what each one does.
| Tool | What it does here |
|---|
| Python | The programming language the app is written in |
| Flask | Turns the Python code into a small website you open in your browser |
| Ollama | Downloads AI models and runs them on your own computer |
| Llama 3 and Llama 3.1 | Free AI models from Meta. Llama 3.1 answers questions, and Llama 3 checks them for safety |
| NVIDIA NeMo Guardrails | A toolkit that asks an AI model whether a message is safe |
| JavaScript, HTML, and CSS | The chat screen in your browser |
| Git | Downloads this project. Optional, since you can download a ZIP instead |
Before you start
You need three things. Each takes a few minutes to install.
1. Python 3.10 or newer. Download it from python.org/downloads.
- On Windows, tick "Add python.exe to PATH" on the first installer screen.
- To check that it worked, open a terminal and run
python3 --version. On Windows, run py --version. You should see 3.10 or higher.
2. Ollama. Download it from ollama.com/download, install it, and open it once. On a Mac a small llama icon appears in the menu bar, which means it's running.
3. Space and memory. You'll need about 10 GB of free disk space for the two AI models, and at least 8 GB of RAM. A machine with 16 GB is faster.
What's a terminal? It's a window where you type commands.
- On a Mac, press Cmd + Space, type Terminal, and press Enter.
- On Windows, press the Start button, type PowerShell, and press Enter.
- On Linux, press Ctrl + Alt + T.
Step-by-step setup
Step 1: Download the project
Option A, with Git:
git clone https://github.com/bayegaspard/responsible-ai-healthcare-guardrails.git
cd responsible-ai-healthcare-guardrails
Option B, without Git: On this GitHub page, click the green Code button and choose Download ZIP. Unzip the file, open a terminal, and move into the folder. For example:
cd Downloads/responsible-ai-healthcare-guardrails-main
Step 2: Run the setup, one time only
This step installs the Python packages, downloads the two AI models, and checks that everything works. The first time takes 10 to 30 minutes, mostly to download the models.
On a Mac or Linux:
./scripts/setup.sh
On Windows:
powershell -ExecutionPolicy Bypass -File scripts\setup.ps1
When it finishes you'll see "Setup complete."
Step 3: Start the demo
On a Mac or Linux:
./scripts/run.sh
On Windows:
powershell -ExecutionPolicy Bypass -File scripts\run.ps1
You'll see "Demo running at http://127.0.0.1:5050". Leave this terminal window open.
Step 4: Open it in your browser
Go to http://127.0.0.1:5050. Type a question and press Enter.
To stop the demo, go back to the terminal and press Ctrl + C. Next time, you only need Step 3.
Try it: three examples
Send each message with guardrails On. Then press the up arrow key to bring the message back, flip the switch to Off, and send it again. Compare the two answers.
This is the beginning of the README. Read the rest at the original. Read the full README ↗
