Du POC à l'Intégration : Comment TheFork a réussi son IA

Go behind the scenes of the creation of "Ask The Fork," The Fork's new semantic search engine using natural language. Hosted by Capucine (Senior PM at The Fork) and Simran (Senior Product Designer at Thiga), this talk sheds light on the concrete challenges, setbacks, and key lessons learned from a large-scale AI project (52,000 restaurants, 11 countries) that transformed a simple Innovation Week prototype into an interface used by millions of people. 📌 Key Takeaways: • The chatbot illusion: Thinking that every AI interface must be a conversational agent is a trap. Finding a restaurant is a 10-second task: the user needs an engine that understands their intent (e.g., searching for a specific "butter chicken" via menus and reviews), not a conversation. The interface must serve the Job-To-Be-Done. • Isolation has its limits (The transition to All-in-One Search): To protect the legacy product during beta, the AI ​​functionality was initially isolated, capping adoption at 3.7%. The true value was unlocked by merging the technologies into a single, universal search bar on the homepage. • The Evidence Layer: Explainability is the cornerstone of trust. Users don't doubt the AI's results; they doubt its reasoning. Displaying verbatim excerpts from verified reviews or menus that justify the recommendation is essential to overcome these hesitations. • From Keyword Matching to Embeddings: AI approximations (like confusing a biryani rice dish with paella) cannot be fixed with simple quick engineering. To ensure the reliability of semantic search, it was necessary to vectorize and model entire sentences in a robust knowledge base (Restaurant Knowledge Base). 🎤 What is La Product Conf? https://www.laproductconf.com/ 💻 What is Thiga? https://www.thiga.co/en/ 📰 Discover our media: https://www.media.thiga.co/ 🎓 Our Thiga Academy training programs: https://www.academy.thiga.co/en/ 🔵 Follow us on LinkedIn: /la-product-conference