About This Demo
Production-ready agentic AI system built with RAG architecture, tool calling, and hallucination prevention.
Try These Queries
- "Contact du restaurant le plus proche de Meudon" (geolocation + fuzzy matching)
- "Recommandez-moi un plat végétarien" (multi-step agentic reasoning)
- "Quelle différence entre Pho et Bun Bo ?" (semantic search)
- "Quels restaurants dans le 91 ?" (department filtering)
Multilingual support: Try queries in French, English, Spanish, or any language
Production Features
- Anti-Hallucination: If info is missing from KB, agent responds "Cette information n'est pas disponible sur le site" (never invents data)
- Geolocation + Fuzzy Matching: Handles "Ivry sur Seine" vs "Ivry-sur-Seine", finds nearest restaurant via Nominatim API
- Multilingual: Auto-detects and responds in query language
- Clickable Links: Preserved HTML from scraped content
- Markdown-Free: Plain text responses (no formatting artifacts)
Agentic RAG Architecture
- 9 Agentic Tools: Multi-step reasoning with parallel execution (search, filter, recommend, geolocation...)
- 4-Layer Validation: Anti-hallucination system (restaurants/schedules/prices/departments)
- FAISS IndexFlatIP: Vector similarity on 20 restaurants + 32 menu items (5-10ms latency, ~500ms avg total)
- Zero Hardcoded Data: Single source of truth from knowledge base (100% RAG)
- Stack: FastAPI + OpenAI GPT-4o-mini (temp=0.1) + text-embedding-ada-002 | Deployed on Render.com with UptimeRobot 24/7
- Automated CI/CD: Weekly web scraping + embeddings rebuild via GitHub Actions