Les Clefs de Versailles

Team Lumière - An intelligent agentic solution for personalized visitor itineraries at the Palace of Versailles.

🎯 The Challenge

Visiting Versailles can be overwhelming:

• Complex palace structure with interconnected gardens and hundreds of rooms
• Information overload and difficulty planning for time constraints
• Weather conditions, crowd management, and diverse personal interests

Our Solution: An AI-powered system that plans custom itineraries through simple natural language conversation, eliminating complexity.

🏗️ Architecture Overview

Frontend

Modern Chat Interface built with React, TypeScript, Vite, and Axios

• 💬 Intuitive & Responsive Design: Clean interface that works on any device
• ⚡ Real-Time Feedback: Live connection status and animated typing indicators
• 🔍 Agent Transparency: Visual badges showing which tools were used (RAG, Weather, Travel, Web Search)
• 🎙️ Voice Support: Natural speech input in French/English for hands-free interaction
• 💾 Persistent Conversations: Unique conversation IDs for resuming sessions
• 🎯 Easy Onboarding: Pre-defined query buttons to discover capabilities

Backend

Modular Agentic Architecture with specialized tools for clean separation of concerns

1. Agent - LLM with Tool Calling

• Tech: OpenAI GPT-4.1 / Mistral Medium 3.1 / LangChain
• Two Modes:
  • One-Shot: Plans complete visits without additional questions (API test mode)
  • Conversational: Asks 2-3 questions to infer user intent and preferences
• Grounded in Versailles knowledge base for factual accuracy

2. RAG Tool - Knowledge Base Search

• Tech: Mistral Embeddings / ChromaDB / LangChain
• 5,799 searchable chunks from 343 official documents
• Sub-second response times with source attribution
• Factual questions about locations, schedules, tickets, and services

3. Weather Tool - Forecast Integration

• Tech: OpenWeather API
• Weather-dependent itinerary planning and recommendations
• Invoked when visit dates are mentioned

4. Travel Tool - Transportation Planner

• Tech: Google Maps API
• Detailed directions with public transit, driving, walking, and parking options
• Multi-modal transport comparisons

5. Web Search Tool - Current Information

• Tech: Linkup API
• Up-to-date information complementing the knowledge base
• Uses keywords from user messages

📊 Data Processing Pipeline

JSONL Refinement (4,700+ entries)

1. Recursive Text Extraction: Traversed nested JSON structures
2. Content Filtering: Removed fragments under 7 words, preserved headings
3. Text Cleaning: Stripped URLs, normalized whitespace
4. Smart List Handling: Joined items with semicolons to preserve context
5. Deduplication: Eliminated duplicates while maintaining order
6. Final Formatting: Created {"url": "text"} structure optimized for embeddings

PDF Processing with Docling

• Layout Detection: ML models identify structure, headers, paragraphs
• Table Extraction: Specialized model preserves data relationships
• Structure Preservation: Maintained folder hierarchy for navigation
• Result: All PDFs converted to clean Markdown for RAG chunking

RAG Knowledge Indexing

1. Raw Data Processing: 343 documents from chateauversailles.fr
2. Intelligent Chunking: 2000-6000 character semantic chunks with 200-char overlap
3. Vector Embedding: Mistral AI converts chunks to numerical vectors
4. Persistent Storage: ChromaDB stores 5,799 chunks with metadata and URLs

RAG Query Processing

1. User question → 2. Vector embedding → 3. Similarity search (top 5 from 5,799) → 4. Context assembly → 5. LLM generates answer with citations

Performance: Sub-second responses with comprehensive source attribution

🚀 Getting Started

Installation

pip install -r requirements.txt

Environment Variables

Create a .env file with:

OPENAI_API_KEY=your_key
MISTRAL_API_KEY=your_key
OPENWEATHER_API_KEY=your_key
GOOGLE_MAPS_API_KEY=your_key
GEMINI_API_KEY=your_key

Running the Backend

python main.py

The backend server will start and listen for requests from the frontend.

Running the Frontend

cd frontend
npm install
npm run dev

Testing the Agent (CLI)

For command-line testing:

python agents/core/simple_agent.py

📁 Project Structure

LumiereVersailles/
├── agents/
│   ├── core/           # Agent implementation
│   ├── prompts/        # Modular prompt files
│   ├── tools/          # RAG, Weather, Travel, Web Search tools
│   └── config/         # Settings and configuration
├── data/
│   ├── rag_data/       # Processed Markdown files
│   └── other/          # Original PDFs and JSONL
├── frontend/           # React + TypeScript chat interface
├── chroma_db/          # Vector database storage
├── fill_db.py          # Build RAG embeddings index
└── ask.py              # Test RAG queries

🎯 Key Features

• 5,799 Knowledge Chunks: Comprehensive official Versailles information
• Sub-second Responses: Real-time personalized planning
• Source Attribution: Verified, trustworthy answers with URLs
• Bilingual Support: French and English interface
• Voice Input: Hands-free interaction while exploring
• Two Operating Modes: API test (one-shot) and conversational

🏆 Team Lumière

Versailles Hackathon 2025

Transforming visitor experiences through intelligent, personalized itinerary planning.