An assistant that actually knows the catalogue.
An invented hardware shop with 193,000 products, and an assistant reading the same data the page does. Not a video.
Loading the demo …
Phone screens are the wrong shape for a shop and a chat panel at once. The demo opens full-screen instead.
Open the demoThree questions a search box cannot answer.
A search box matches words. None of the questions below are about words — each one needs the machine to understand what a product is and what it connects to.
Reads the socket off the product, finds matching bulbs, and excludes anything over the fixture's wattage limit.
Checks the IP rating and the manufacturer's stated use against the bathroom zones in DIN VDE 0100-701 — and says no when the answer is no.
Asks the questions that actually change the answer, then builds a bill of materials with quantities.
The shop, brands, products and prices in this demo are invented. The catalogue is synthetic and contains no customer data. To prevent abuse, the number of questions per visitor is limited.
The catalogue is the model, not the prompt.
Nothing here is the language model being clever. The parts that must be right are computed from the data, and the model is only allowed to speak about what it has actually read.
Sockets, wattages, IP ratings and mount types are pulled out of the product text into fields that can be compared. That is the unglamorous half of the work, and the half everything else rests on.
Whether two parts go together is decided by a rule over those fields, including numeric limits. The model cannot talk its way past a wattage ceiling.
When the answer depends on something only you know, it asks a short question instead of picking for you. When the data does not answer, it says so rather than inventing a plausible product.
See this on your own data.
We build the same mechanism on your catalogue, your datasheets, your terms. One conversation is enough to work out whether it is worth it for you.
Book a scope call Or write directly: hello@deepstill.de