A search engine that shows its working

The problem

A buyer sourcing a DN50 stainless steel ball valve for a chilled-water line knows exactly what they need. What they get from most catalogues is a keyword box, forty facets, and a results list ordered by something nobody explains. The specification — the part that decides whether the product is usable at all — sits in a PDF datasheet three clicks away.

So the buyer emails five suppliers, describes the requirement in prose five times, and waits. The catalogue was never the bottleneck. The translation between how a requirement is stated and how a product is indexed was.

What we built instead

Every product carries typed specifications — material, connection type, nominal diameter, pressure rating, ingress protection — governed by a central registry rather than free text. That registry is what makes the rest possible: it drives the facets, the comparison table, the vocabulary the assistant understands, and the weights used to rank.

You describe the requirement in your own words. That description is parsed into a structured intent: category, specifications, budget, application, and anything you have ruled out. The catalogue is filtered on the hard constraints, and everything that survives is scored against eight weighted components.

The language model never decides anything

This is the part worth being precise about, because it is unusual. When a language model is configured, it does exactly two jobs: it turns your sentence into that structured intent, and it rephrases results the engine has already computed.

It does not filter the catalogue. It does not rank. It does not price anything, and it never produces the match percentage. Those come from deterministic code — src/server/catalog/scoring.ts — that you can read. Turn the model off entirely and the same query returns the same products in the same order with the same scores. Only the prose gets terser.

The reasoning is simple. A number a model generated is a number that can be wrong in ways nobody can audit, and a buyer is going to commit real money on the strength of that figure. It has to be reproducible.

Why the score is never 100%

Match percentages are banded between 42 and 97 — not because the arithmetic cannot reach 100, but because a perfect match is a claim nobody can stand behind. There is always a requirement the buyer did not state. The panel beside each result lists what matched, what did not, and what the catalogue has no data for. A recommendation surface that hides its misses is marketing.

What happens next

There is no cart. Industrial supply is quoted, not checked out: price depends on quantity, delivery, GST treatment and the relationship. So a shortlist becomes a quotation request, and the platform carries it through pricing and response with the specification still attached — which is the part that usually gets lost in an email thread.

Read the frequently asked questions, or try it directly. The assistant needs no account.