Versailles Concierge — Les 4 MousquetAIres

“Where should I start? Will it rain? Can I still see the fountains?”
During the Datacraft × Château de Versailles hackathon, we kept hearing these questions from visitors and staff alike. Versailles Concierge is the assistant we built to answer them—politely, quickly, and in their own language—then turn those answers into a live, navigable plan.
➡️ Live demo: vers-ai-lles.vercel.app

Why we built itCopy link to section

Versailles is magical and massive: timed tickets, long queues, 800 ha to cross, shifting weather, and crowd spikes that can derail even a careful itinerary. Our goal felt obvious once we walked the grounds:

• Welcome anyone, in their language, with fast, grounded answers.
• Transform a chat into a plan—timeline, map, and route options.
• Adapt in real time to weather, timing, and where the visitor actually is.

A small personal note: the “aha” moment came when a staff member asked us, in passing, “If it rains after 3pm, would you reorder the day for a family with a stroller?” That single constraint captured the whole product: context-aware guidance that respects the reality on site.

What visitors seeCopy link to section

Our main interface (web/) is a Vite + React SPA tuned for kiosks and phones:

• Conversational chat with pre-seeded prompts and bilingual copy in constants/translations.
• Plan board that turns the agent’s payload into a crisp timeline of highlights and notes.
• Interactive map (Leaflet) for curated routes, point-to-point directions, or a multi-stop demo.
• “Meta” hints (weather, likely crowding) via a MetaSummary that blends live APIs with time-of-day heuristics.
• Local persistence for session state, language, history, and last itinerary.

We first validated flows in a Streamlit prototype (frontend/), then ported the experience to React once the pieces felt right.

The concierge under the hoodCopy link to section

The backend (backend/app.py) is FastAPI orchestrating a LangChain + LangGraph agent:

• Chat runs on mistral-medium with structured tool calls and scoped memory.
• A LangGraph state machine (backend/setup_graph.py) routes between Q&A, plan creation, and tool execution.
• Memory is per session, so the assistant actually remembers visitor context.
• Clean endpoints: / (full agent), /chat (stateless), and /tools/* for weather, routing, GPS sharing, and plan recalculation.

Knowledge & tools that stay groundedCopy link to section

• A curated corpus (backend/data/tips_embedded.json), embedded with mistral-embed and retrieved by cosine similarity.
• Model Context Protocol bridges (mcp_servers/) for OpenWeather / Open-Meteo and OpenStreetMap + OSRM routing.
  The LLM asks; the bridges fetch; the agent composes a safe, auditable reply.
• /tools/recalc_plan can rebuild an itinerary on the fly—merging shared GPS, default waypoints, and constraints—then returns GeoJSON + human-readable steps.

Shipping fast without chaosCopy link to section

• docker-compose.yaml spins up FastAPI, MCP servers, and optional Streamlit in one go.
• Environment variables (see README.md) centralize keys for Mistral, weather, and maps.
• We kept clear boundaries: Python backend, React web app, and MCP micro-services each with their own deps and hot-reload loops.

The 48-hour sprintCopy link to section

1. Walk the journey with Versailles staff; collect the pivotal questions and choke points.
2. Prototype in Streamlit to test tone and flow; harden the UX in React once it clicked.
3. Assemble the agent with stateful tool-calling, RAG, and multilingual prompts that degrade gracefully when data is missing.
4. Wire weather & routing MCPs so “Should I start with the gardens?” yields actionable, timed advice.
5. Demo end-to-end: day plans, live routes, and contextual tips—in English or French—from a single chat window.

Honestly, the best part was the corridor testing: a guard asked us for a path that avoided stairs at closing time; a family wanted to front-load the gardens before a forecasted shower; a couple had only 90 minutes. Each mini-story nudged the product toward practical, empathetic decisions—and away from generic chat.

Deployed demoCopy link to section

• Frontend: Vercel — vers-ai-lles.vercel.app
• Backend: Render (FastAPI + MCP services)

If you’ve ever stood under the gilded gates wondering “what’s the smartest next step right now?”—this is our answer. Built in two days, shaped by real conversations, and ready to guide the next visitor who asks.