01
The catalogue expected customers to know the product name
Technical buyers usually arrive with a requirement, not a perfectly formed catalogue query.
A conventional search works when the words typed resemble the product title. It is less helpful when a customer asks for a 2-inch brass ball valve with a lever, a WRAS-approved valve for drinking water, or the stainless equivalent of something they have already found. Those requests combine engineering attributes, synonyms and follow-up context in ways a keyword box does not naturally resolve.
The goal was not an open-ended support bot. It was a narrower product-finding layer that could understand the request, stay inside the published catalogue and hand the customer back something they could inspect or buy.
02
We made every product searchable as a set of engineering constraints
Each catalogue item became a compact passage of the attributes that actually distinguish it.
An offline indexing pass embeds product name, category, size, material, connection, pressure rating and WRAS status. The 1,126-product index is compressed to 1.5 MB and ships inside the existing Vercel function, where a full cosine scan takes less than a millisecond. At this catalogue size, a vector database would have added moving parts without improving the answer.
Cosine similarity finds the general meaning; a small lexical re-rank enforces explicit requirements such as brass, DN80, flanged or WRAS-approved. The best eight products go to a compact language model, which can explain the result but can only recommend from that retrieved set. Product cards are then attached from catalogue SKUs actually named in the reply.
03
The product finder went live without adding another platform
The widget runs on every page, remembers the conversation across navigation and drops customers onto a real product page.
In a small pre-launch staff test built around representative product-finding jobs, median time to a relevant product fell by 72%. All six retrieval benchmark shapes returned the right material, size and category within the top five, including follow-ups such as “what about in stainless?”.
The request path stays inside the existing Express-on-Vercel application. Inputs are capped and rate-limited, supplier and cost fields never enter the prompt, model output is escaped, and 29 API and browser checks cover realistic searches, context, prompt injection, data isolation and the complete reply-to-product-card flow. A typical message costs around a tenth of a penny.
Can customers describe the product before they can name it?
If your catalogue is technical, large or full of near-matches, a grounded product finder can shorten the route from requirement to product without replacing the systems already underneath it.