bookmi.
Try free
← Back to blog
Tech2026-01-05· 11 min· Bookmi Tech Team

RAG for Restaurants: Optimizing Menu, FAQ & Business Info Retrieval

Technical details on pgvector + Hybrid Search implementation, chunking strategies, and real-time updates.

RAG for Restaurants

"Is the seasonal white peach parfait still available?"

A customer asked via LINE, and our AI immediately replied: "Yes, we have it!"

The problem? The white peach parfait had ended 3 days ago.

When the store manager called us in frustration, we realized a critical flaw. AI possesses vast "general knowledge," but knows nothing about "what's happening in the store today."

AI's "Amnesia"

ChatGPT or Claude can answer anything included in their training data. But today's menu? This week's business hours? The item that just sold out?—These don't exist in the training data.

We needed a way to teach AI "the store's present."

Injecting Memories

What we built is RAG (Retrieval Augmented Generation)—a mechanism that "injects" relevant store information just before the AI generates a response.

When a customer asks about "white peach parfait," before letting the AI answer, we first search the database. We find information related to "white peach parfait"—price, availability period, remaining stock—and pass that information along with the question to the AI.

The AI generates its response based on "injected memories."

Finding the Right Memory in 0.1 Seconds

But search has its challenges. A customer might ask about "peach dessert" or "peach parfait." They're all referring to the same "white peach parfait."

Keyword search alone can't handle this. We introduced vector search that can search by "meaning." By comparing the meaning of the question with the meaning of menu items, we find the most relevant information within 0.1 seconds.

Reflecting Changes from 1 Second Ago

Even more important is information freshness. The moment a manager clicks "White Peach Parfait: Sold Out" on the management screen, that information must be reflected in the AI's "memory."

We built a system where menu and inventory changes sync to the vector database in real-time. From change to reflection: less than 1 second on average.

An AI That Can Say "No"

Since that white peach parfait incident, the AI can now accurately answer: "We apologize, the white peach parfait was a seasonal item and ended 3 days ago."

And it adds a suggestion: "As an alternative, we currently offer a kyoho grape parfait."

By giving AI accurate memories, we can deliver accurate answers to customers. That's the purpose of our RAG.

Answering calls, replying to reviews, following up

AI can take over most of the work that eats your time. Try it for 30 days, no credit card.

Start 30-day free trial →Read other articles →