One vocabulary for 34,000 products
KVR Audio's product database had grown past 34 000 plugins, instruments, and soundware. Classification was whatever a developer typed into one tags field. We wrote a controlled vocabulary, mapped the legacy tags onto it, and classified the live catalog so search and browse follow governed terms, not tag soup.
A catalog classified by whoever typed last.
Format and platform tags on KVR were already reliable. Functional classification was not. One flat tags field was asked to hold what a product is, what it does, what it sounds like, who it is for, and how it is licensed — all at once.
Developers supplied free-text tags. A reverb in the live catalog would pick up compressor, EQ, and vintage amp next to its actual reverb tags. Search and browse followed the noise. A small team could not retag 34 000 products by hand.
Facets, a closed list, and a human gate.
We did not ask a model to invent a taxonomy. We wrote one.
Classification split into independent facets: product class, category, subcategory, then character, use, and genre where they apply. Format, platform, and licensing stayed on the data that already worked. Each product takes exactly one class and one primary category, from a closed list. Custom tags stay attached to the product; they no longer drive navigation.
The catalog pass is mostly automatic. An alias map sends common legacy spellings to a canonical term with no model involved. What the map cannot place goes to a classifier that can only pick from the approved list or say it is unsure. Unsure cases queue for a person. Developer submission forms now pick from the same list, so the mess does not grow back.
Faceted vocabulary
Nine kinds of information stop sharing one field. Terms are singular nouns in Title Case, one concept, one facet. A product is described by picking across the list, never by inventing free text.
Alias map
The common legacy spellings, abbreviations, and near-synonyms resolve to one canonical term. That pass classifies the obvious majority of the catalog with no model and no ambiguity.
Constrained classifier
What the map cannot place is proposed by a model that may only choose from the approved list or mark the row uncertain. It cannot invent a tag.
Keep, map, demote
Legacy custom tags stay on the product and remain searchable, so inbound links and long-tail queries do not break. Filters and navigation are built only from controlled terms.
Fix it at the source
Developer submission and edit forms pick class, category, and subcategories from the same dropdowns. Free-text tags remain optional and secondary, so the catalog does not refill with soup.
The live catalog has a structure now.
The product database — 34 000+ entries — is classified against the controlled list. Search, filters, and category pages run on governed terms. Developer-supplied tags remain searchable; they no longer set the structure.
The vocabulary is owned. New terms go through a request, not a free-text field.