AI Product Concierge — Interactive Diagnosis & Recommendations

Introduction

If you asked me what makes a physical store great, I would say "the service" without hesitation. A veteran salesperson draws out your intended use and preferences from casual conversation, then recommends exactly the right product — with reasons attached.

Now think about an EC site. Finding the right product is left entirely to the customer, armed only with search and category filters. In this article I introduce the overall picture of the AI product concierge we built for an EC site selling motorcycle gear. It interviews customers, hears out their requirements, and follows through to grounded product recommendations.

Recreating in-store service online

What EC sites are missing

EC sites have product information and search. Even so, the complaint "I can't tell which one is right for me" never goes away. Especially for specialized products like helmets and jackets, staring at a spec sheet makes the decision hard.

In a store, the conversation begins with a single line — "What kind of riding do you do?" — and the candidates naturally narrow down. This narrowing through dialog is the experience EC has been missing.

Breaking down what a veteran does

To systematize service, I first broke down what a veteran salesperson actually does. Organized, it splits into three roles.

1. Interview — draw out use case (touring, commuting, etc.), preferences, and budget
2. Match against knowledge — pick candidates from product knowledge and experience
3. Recommend and explain — convey "why this one" with reasons attached

This system maps each of those three to a technical component: dialog design, RAG search, and a visualization UI.

The interactive diagnosis format

When people hear "diagnostic content," the classic image is a questionnaire where you pick options in order. But a questionnaire requires designing every branch in advance and doesn't allow free-form consultation.

So we made it interactive using an LLM. The customer can simply say "I go on long-distance touring on weekends" in their own words, and the AI reads use case, preference, and budget from that. If something is missing, the AI asks about it.

The system at a glance

Three building blocks

The system is built in three layers: a dialog engine, a knowledge base, and a recommendation UI.

The dialog engine converses with the customer while, behind the scenes, it searches the product master and knowledge base and the UI presents the results clearly.

A product master linked to Shopify

The product information behind every recommendation is synced periodically from Shopify's product data. The key is pulling in not just names and prices but variants (color options and sizes) too.

Recommending a size that is out of stock is pointless, and being able to say "this color is available" is what makes it service. On top of that, we built in size-recommendation logic that derives the right size from body dimensions like height and build, compensating for the fact that you can't try things on online.

A knowledge base for grounding

The product master alone can't answer deeper questions like "how waterproof is it?" or "what material is it made of?" So we extracted text from product manuals (PDFs) and web product information, chunked and embedded it, and stored the vectors in Neon Postgres (pgvector).

Before recommending, the AI searches this knowledge base and explains using only what the actual documents say. This RAG (retrieval-augmented generation) setup is the linchpin that prevents plausible-sounding lies.

From dialog to recommendation

Gathering requirements through the interview

When a customer starts the diagnosis, the AI first asks about use case. In the conversation that follows, it records the information it reads from each answer into state, and only asks about the fields still empty.

What I cared about most was not over-asking. Bombarding a customer with questions makes them leave, so we move on to recommendations the moment the minimum required fields are filled.

Grounded recommendations

Once the requirements are in, we narrow candidates from the product master, back up each candidate's features against the knowledge base, and then generate the recommendation text.

Being able to give an explanation with a source — "for touring, this product has the ventilation mechanism described in its manual, so it stays comfortable on long rides" — is RAG's strength. Not letting the AI speak from imagination is especially important in EC service.

Visualization builds conviction

Recommendations are shown not just as text, but with a comparison table of candidate products and a radar chart of their characteristics. When you can line them up side by side, "why this one" comes across at a glance.

Furthermore, we present related items that pair well with the chosen product (like an intercom for a helmet) as upsells. It recreates that in-store moment of "how about this to go with it?"

Conclusion

I've introduced the overall picture of the AI product concierge. There are three key points.

1. Decomposing in-store service — mapping interview, knowledge matching, and explanation to technical components
2. Grounded recommendations — product master × RAG prevents plausible-sounding lies
3. Visualization — comparison tables and charts build conviction and nudge the purchase

The details of each element are explored in their own articles. For the interview question flow and state management, see "Designing the Interview Dialog and State Management." For building the knowledge base and how RAG works, see "Product Master × RAG for Grounded Recommendations." For the comparison table and chart UI design, see "Visualizing Recommendations with Tables and Charts."