A RAG Chatbot for Saigon Food

npx skills@latest add phucbm/skills --skill rag

food-tour-sg is a chatbot that finds places to eat in Saigon from a list of restaurants, instead of from what the model remembers. The data lives in CSV and Markdown files in the repo; an ingest script embeds them into Pinecone, and every message pulls the closest chunks into the system prompt before the chat model answers. The recipe is split across four skills: rag for the pipeline, pinecone for the store, vercel-ai-gateway for the models, and groq for streaming from Groq directly, which came from OpenWallet's admin. I built it to learn RAG on data I know: where I eat in Saigon.

What I didn't expect

• The vector DB doesn't answer anything. It returns the chunks closest to the question, as text. The chat model never sees Pinecone; it only reads what I paste into the prompt. Once I saw it that way, the two API keys (one for embeddings, one for chat, often different providers) made sense.
• Retrieval goes in the prompt, not in a tool. I search on every message and inject the result, so the model doesn't decide whether to look. It is one round trip less and the model can't skip it.
• Filters cost recall. A metadata filter (a dish, a district) runs before ranking, so a top 5 that was fine without filters returns too little with one. I raise K to 10-30 when a filter is on.
• Vietnamese names need tokens. People type "com tam" and "quan 1" without accents, which doesn't embed close to "cơm tấm" and "Quận 1". Each chunk stores normalised tokens and the query matches them with $in; no match falls back to plain semantic search.

Re-ingesting

A GitHub Action re-runs the ingest when anything under data/ changes on main, so a new place is in the bot after a push. Re-embedding every chunk is the expensive step, so it only runs when the data changes.

Models