ai_shop_assistant

The problem

An AI model with no way to look things up will happily invent a product, a price or a link. For shopping that is a real problem: a made-up price is worse than no answer.

What I built

A Rails chat app that gives the model one tool, SearchProducts, and lets it decide when to use it. When a question needs real product data, the model asks for the tool, the app searches Shopify’s live Global Catalog, and the model answers from what came back: up to three products, each reduced to a title, a price and a URL.

It is the same tool calling pattern behind the shopping features in ChatGPT, Perplexity and Copilot, built from scratch in Ruby.

How it works

1. Save and enqueue. The controller saves the message and hands the slow model call to a background job, so the web request returns at once.
2. Run the model in a job. ChatResponseJob on Solid Queue builds a RubyLLM chat with the SearchProducts tool attached.
3. Let the model decide. If it asks for a search, the browser shows a “searching” status line.
4. Search the live catalog. The tool gets an OAuth 2.0 client-credentials token and sends a JSON-RPC tools/call to Shopify’s catalog MCP endpoint.
5. Stream the reply. The saved message broadcasts itself to the browser with a Turbo Stream.

What I verified

• 14 RSpec specs with RubyLLM and HTTP stubbed, including the conversation-memory case (“My name is Alex”, then “What is my name?”).
• End to end against Shopify’s live Global Catalog: the model chose when to search, and replies came back with real products, prices and links, streamed to the browser.
• CI on every push: RuboCop, Brakeman and bundler-audit. The specs run locally.

What broke, and what I learned

Replies never reached the browser. The job and the web server are separate processes, and Action Cable’s default async adapter only broadcasts inside one process. I switched to Solid Cable, which passes broadcasts through a shared database, with no Redis needed.

The assistant forgot the conversation. Each job builds a fresh chat, so on the second message the model had no memory of the first. The fix replays saved messages into the new chat, and a spec covers it (“My name is Alex”, then “What is my name?”).

An MCP gem crashed on the handshake. ruby_llm-mcp failed on Shopify’s response, so I filed ruby_llm-mcp#155 and called the endpoint directly with a small hand-written JSON-RPC client over Net::HTTP.

Approach: key decisions

• One tool, not many. The point was to see the model choose when to search, so the tool surface stays minimal.
• MongoDB for chats. Chats and messages are Mongoid documents, so RubyLLM’s ActiveRecord-only acts_as_chat could not be used and persistence is written by hand.
• Stubbed specs. All 14 specs stub RubyLLM and HTTP, so they are fast and never spend API quota. Live behavior is checked by running the app.

Status

Works end to end: chat, live catalog search and streamed replies. It is a learning and portfolio project. It only searches (no cart or checkout) and returns at most three products per search.

Related services

Need something like this in your own product? See AI engineering and integration and Shopify, BigCommerce and e-commerce.