wagami-rag-en

wagami-rag-en is an agentic Retrieval-Augmented Generation chatbot for a restaurant menu. Guests ask about dishes, prices, per-serving nutrition and allergens, and receive answers grounded entirely in a curated knowledge base rather than a language model's own knowledge. The system is English-only and includes a FastAPI service, a LangGraph agent, a guest-facing chat page, a Telegram bot that talks to the same agent, and tooling that loads the knowledge base into Weaviate Cloud. A guest asks something like "which vegan ramen has the fewest calories?" or "does the katsu curry contain gluten?". The agent turns the question into a hybrid (vector + keyword) search over a Weaviate Cloud collection, reranks the hits, and an LLM composes the answer from only those rows. Every dish the answer contains is shown as a card with its photo, and only those dishes — never a card left over from an earlier answer. When the knowledge base doesn't hold the answer, the agent says so rather than inventing one. Not every message needs a search. One LLM call reads the message and decides how it is handled: greetings and off-topic messages get fixed replies, menu browse requests list groups, categories or items straight from the corpus, and dish, ingredient, price, allergy, or filtered questions go through hybrid search, rerank, and a grounded answer. The same agent answers on two surfaces: the web chat page and the Telegram bot. Both run the same chat turn, under the same token-budget, conversation-length and concurrency caps.