Demo-AI-Conversation

This repository contains the backend code for the Cornelius AI Project: Talk2Care. Talk2Care is designed to be used for research in an educational settings, too see how well AI can be used as an educational tool. For this case it is meant too simulate real-world medical scenarios. The AI acts as a virtual patient, allowing users to practice their communication skills before engaging with real patients.

While the backend is primarily designed to work with a frontend, it can also function independently. You can interact with the backend directly using the client.py file, which allows you to send requests to the backend for processing.

System Requirements

The system requirements may vary depending on the models being used. Below are the minimum requirements necessary to run the project:

• Docker: Required to containerize the backend services.
• Python version 3 or later: The project is built with Python 3.
• 16 GB RAM (minimum): Docker and AI models have a lot of memory demand.
• Nvidia GPU (optional but recommended): The models used in development are boosted by the GPU. However if Nvidia is not available it will default to CPU.

Operating Systems:

• Linux (recommended)
• macOS
• Windows (with Docker Desktop)

Installation

1. Clone the repository:

   git clone https://github.com/jrcz-data-science-lab/demo-ai-conversation.git
   cd demo-ai-conversation

2. Build the backend using Docker:

   docker compose up --build

3. [OPTIONAL] Run the client:

   python client.py

NOTE: The final build includes large models. If you prefer a lightweight setup, it's recommended to replace the models before running the Docker commands.

Once the backend is running locally, you can access it at http://localhost:8000/general this is the main entry point. If you are only working with the backend, you can use the client.py file to send requests to the backend.

Deployment Environment

The current build is meant for server deployment. Therefore, the code is written to use larger models when running in this environment. All models are run locally and do not use any remote AI services.

If you run the server in it's deployed environment you will need to call it from the frontend application or run the client.py script remotely.

Commands

sudo docker compose up will start the server

sudo docker compose up --build will build the application and the models

sudo docker compose down will remove all containers. This is recommended if you are replacing the models. After running this command you need to rebuild the project before starting it again.

Changing Models

Speech-To-Text

STT Component in faster-whisper.py file on line 11 you can change the large-turbo-v3 this specifies the models for faster whisper refer to the repository for a list of models and there specifications: https://github.com/SYSTRAN/faster-whisper

model = WhisperModel("large-v3-turbo", compute_type="int8")

Ollama

Ollama Component you will need to change two files in Ollama's folder go to the entrypoint.sh file and change the variable in line 3 or 4 for the models that you like to use:

CONVERSATION_MODEL="gemma3:27b"
FEEDBACK_MODEL="qwen3:32b"

These two models have different tasks. If you change the conversation model you need to go to the manager folder and go to app.py in line 163 you will need to change the model in the parameters here too:

ollama_response = requests.post(
  OLLAMA_URL,
  json={"prompt": prompt_text, "model": "gemma3:27b", "stream": False, "think": False}
)

If you change the feedback model you have to go the same file and instead go to line 214 change the parameter of the model you like to use:

ollama_response = requests.post(
  OLLAMA_URL,
  json={"prompt": feedback_prompt, "model": "qwen3:32b", "stream": False, "think": False}
)

For a list of models you can use the Ollama library provided in this link: https://ollama.com/library

Text-To-Speech

TTS Component in tts.py at line 22 you can change xtts_v2 to the model you like to use here: https://github.com/coqui-ai/TTS

tts = TTS("tts_models/multilingual/multi-dataset/xtts_v2").to(device)