How ShopAssist works
Tools, grounding, UI blocks and confirmation cards: how the assistant finds answers in your store data, and what it can't do.
ShopAssist is an AI assistant built on a large language model. This page explains how it reaches an answer, so you can supervise it well (AI literacy, EU AI Act Article 4).
Tools, not memory
The model doesn't know your store. For every question it calls tools that ShopAssist runs on your data: product search, product details, stock per location, store hours, knowledge search, order lookup for verified shoppers, cart and checkout actions. The model writes the answer from the tool results.
Grounding
After the model writes a reply, ShopAssist checks it before the shopper sees it: every price, percentage, stock quantity, duration and time in the text must appear in a tool result or the conversation. A reply that states a number from nowhere is corrected or replaced. Prices and stock are never taken from the knowledge base or from the model.
UI blocks
Products, comparisons, store availability, order status and forms appear as structured cards (UI blocks) rendered by the widget from your data, not as text the model typed. A product card always shows the price and stock from the catalog, formatted for the shopper's language and currency.
Confirmation cards
The assistant can propose actions (cancel an order, start a return). It can never execute them by itself: the shopper confirms on a card, and ShopAssist re-checks the order with your platform first. A shopper can always withdraw from a distance purchase; the assistant never refuses or discourages that.
Disclosure
Shoppers are told they are talking to an AI, every time, and can ask for a person. You can't remove the disclosure (AI disclosure).
Limitations
- The model can misunderstand a question or choose the wrong product. Product cards and order data are always correct; the surrounding text can be imperfect.
- It only knows what is in your catalog, knowledge base and order data. Missing information becomes a knowledge gap.
- It gives no medical, legal or financial advice, and declines topics you block.
- Image understanding (visual search) describes photos; it doesn't identify people.
See supervising your assistant for how to keep quality high.
Last updated on