Find real tags by inventing plausible ones

A blog with 1,856 tags is an awkward classification problem: the list alone can overwhelm a prompt. Doug Turnbull’s approach, highlighted by Simon Willison, starts by letting the model suggest labels without seeing the full vocabulary. Embeddings then match those suggestions to the nearest real categories.

A few examples can show the model what a category looks like without supplying every possible answer. The invented label becomes a search query; the selected label still comes from the existing catalog.