For Velgrina’s US store, I built a streaming product-help chat powered by GPT-4o-mini. It was one of two production AI features in the project, alongside a separate kitchen visualization feature. The chat addressed a specific product need: helping a visitor ask questions while considering kitchenware.
That context matters. A shopping assistant is part of a decision that can lead to an order. The standard for a useful interaction is whether the visitor receives relevant, understandable guidance with a clear next step.
Product questions contain different kinds of uncertainty
Some questions are descriptive: what a product is made from, how it is used, or how two options differ. Others depend on information that changes, such as stock or delivery availability. A third group depends on the customer’s circumstances, including available space or the item they want to use alongside it.
I use that distinction when designing and reviewing product assistance. A material description can be checked against catalog content. A delivery promise needs a current operational source. A question about fit may require a measurement from the customer before any recommendation is useful.
For example, “Will this fit my sink?” is incomplete without knowing which product and which sink. Asking one relevant follow-up question is better than producing a confident answer from the product name. This is a design principle for commerce assistance, not a claim that a language model can infer physical dimensions from a vague message.
Streaming changes the interaction
The implemented chat streams its response. That gives the interface a way to show progress while text is being generated, rather than waiting for a complete answer before changing the screen. It also creates additional states that deserve attention.
A review should distinguish a request being accepted, a response arriving, a response finishing and a response stopping early. If the connection ends halfway through a sentence, the interface should not silently present that fragment as a completed recommendation.
The customer also needs an understandable way to recover. Repeating a question, choosing another product or returning to the product page should remain possible. The value of a conversational interface drops quickly if it interrupts the shopping journey when something fails.
A useful answer needs somewhere to go
Product assistance works best when it stays connected to the store’s ordinary navigation. A visitor may want to inspect specifications, compare an option, look at images or continue to a product page. The conversation should help them reach those resources instead of becoming a separate destination with its own version of product truth.
That is also how I think about scope. An assistant can explain an available specification without making a suitability guarantee. It can point toward information without inventing a policy. When the necessary detail is unavailable, saying what is missing is a useful outcome.
These boundaries are especially relevant for AI features placed beside product photography. Visual confidence and fluent language can make an answer feel more authoritative than its evidence supports. The product experience should make uncertainty understandable in ordinary language.
Review the shopping task
My evaluation checklist for this kind of feature starts with actual visitor tasks: choosing between two items, understanding a material, checking a dimension and recovering from an unclear answer. I also include unrelated questions and requests for information the store does not provide.
Those cases give a more useful review than asking whether the assistant sounds friendly. They expose where the interface needs clarification, where catalog information is insufficient and where the assistant should send someone back to an authoritative page.
The Velgrina implementation gave me another practical AI integration beyond a standalone chat demo. It placed generation inside an existing commerce experience, where response behavior, navigation and product information all affect whether the feature is useful. A successful answer should help the visitor make a better-informed decision, with the limits of that answer intact.
Updated 26 September 2026.