LuminaSwiss is a secure, private, RAG-driven knowledge management system designed to transform internal documentation into intelligent training modules. The project was realized during the Devpost genAI Zurich 2026 Hackathon, it is not intended for production.

Lumina Swiss is a privacy-first Learning Management System designed for organizations that want full control over their data. Built with a "Privacy-First" architecture, it detects and anonymizes PII (Personally Identifiable Information) on the client side and send data to a personalized cloud infrastructure ensuring sensitive information never leaves the organization unprotected. The platform leverages this secure knowledge base to generate personalized training courses tailored to each organization's content and needs. Built entirely on Swiss infrastructure — with API calls routed through Infomaniak and end-to-end encrypted communications — Lumina Swiss guarantees data sovereignty and regulatory compliance. The modular architecture is designed to scale beyond a single organization.

flowchart TD
    subgraph Client["🖥️ Client (Browser)"]
        A[Interface TypeScript / React]
        B[Local Anonymizer\nPII + NER + Tokenizer]
        C[Local Token Storage]
        D[Upload anonymized documents]
        E[Prompt — Personalized course request]
    end

    subgraph Backend["⚙️ Backend (Raspberry Pi / Cloud)"]
        F[API Endpoint]
        G[RAG Ingestion\n bge-m3 + Qdrant]
        H[RAG Engine\n Query enrichment]
        I[Cloud LLM — Infomaniak\n Course generation / Responses]
        J[Response returned to client]
    end

    A --> B
    B --> C
    C --> D
    C --> E
    D -->|HTTPS anonymized doc| F
    E -->|HTTPS prompt| F
    F --> G
    G --> H
    H --> I
    I --> J
    J -->|HTTPS response| A

LuminaSwiss is designed for deployment on local infrastructure and with Swiss base solution LLMs (e.g., Infomaniak Jelastic/Public Cloud).

• Frontend: Static build deployed to a web server or CDN.
• Backend: Python app served via Uvicorn
• Database: Persistent SQLite volume

• Client-Side PII Detection: Uses a local Transformers.js model to detect sensitive data (names, locations, etc.).
• Tokenization: Sensitive data is replaced by tokens (e.g., [[PII_001]]) before upload.
• RAG-First: All LLM queries are grounded in your private, anonymized document base.

• Preferably Linux distro (Ubuntu, Debian...) or WSL
• Node (v22+)

• Debian Bookworm or WSL for Windows users
• Python (v3.11+)
• Docker (v29.0+)
• Port Availability: Ensure ports 8000 (Backend) and 8080 (Frontend) are not being used by other applications.
• Infomaniak LLM API token Key

cd ./server
bash install.sh

cd ./client
npm install
npm run dev  # or : npm run dev:cloud      if you only want to test the client with lumina backend cloud

This guide provides the necessary steps to correctly configure and test the Lumina Swiss system in a local environment or test the client with our deployed backend on the cloud.

If you test the client with Lumina cloud infrastructure you can skip this step. But feel free to reach us on discord to so we can create a user account to access LLM prompt and LMS courses generation. Ensure valid API keys for Infomaniak or OpenAI are provided to enable AI features. An example of the .env below :

cd server && docker compose down -v

Create a .env file in server/

# .env
INFOMANIAK_API_KEY=your_infomaniak_api_key_here
INFOMANIAK_PRODUCT_ID=your_product_id_here
INFOMANIAK_MODEL=llama3  # Available models: mistral3, llama3, mistral3, etc.
CORS_ORIGINS=allowed-frontend-url-origins # default: http://localhost:8080
CADDY_HOST=localhost  # default: localhost, change to it to http://[SERVER-IP] if you are not executing the server locally

Launch server for the first time by launching in server/ :

docker compose up --build -d

The choice of VITE_API_URL is critical for successful communication:

• Same-machine testing: Set VITE_API_URL=http://localhost:8000 (or leave empty).
• Remote local Network testing: Set VITE_API_URL to the machine hosting the server (e.g., http://192.168.1.XX:8000).
• Client test with Lumina cloud infrastracture: Set VITE_API_URL=lumina-swiss.cloud

• Large Files (Up to 20MB): The system is optimized for massive documents. When uploading, look for the real-time progress percentage in the document list.
• Hardware Acceleration: For the fastest PII detection, use a modern browser (Chrome/Edge/Safari) with Hardware Acceleration enabled in settings.

• Generation Time: Complex course generation involves deep RAG analysis and can take 1 to 3 minutes.
• Language: The AI automatically inherits your UI language (French, Italian, German, Romansh).

• Progress Tracking: The quiz progress bar updates specifically when a question is answered. It reaches 100% only upon completion of the final question.

To safely stop all services and preserve resources:

1. Stop the frontend terminal (Ctrl + C).
2. Stop the backend:

cd server && docker compose down

Testing Request: Please test the LLMs prompt and courses generation in either English or French.