What this is

Engineer Doug Turnbull shared a counterintuitive classification trick: instead of asking the LLM to pick from your existing tags, let it freely generate new terms — what the industry calls hallucinate (the model producing content not grounded in its input) — then use vector embeddings (turning text into a string of numbers for similarity math) to locate the closest matches in your existing taxonomy.

He used a furniture e-commerce example: given the query "brown coffee table," let the model invent a few category suggestions (such as "Furniture / Living Room Furniture / Coffee Tables"), then compute similarity between those new tags and your 1,856 real tags, and pick the closest matches.

This approach sidesteps the traditional classification dilemma: a vocabulary too large to fit in the prompt, or one too small to provide adequate coverage.

Industry view

We notice that the one surfacing this is Simon Willison — one of the most influential AI practitioner bloggers in the English-speaking world. Putting it in a prominent spot on his blog is itself an endorsement.

What's more worth attention is the underlying thinking: separate "what the LLM is good at" from "the constraints you need it to obey." The LLM handles creative leaps; vector search pulls it back to reality. This is in line with the RAG (Retrieval-Augmented Generation — retrieve relevant documents first, then let the model answer based on those results) pattern the industry has been championing for the past two years.

But there are dissenting voices. The search-engineering community points out that vector matching itself has errors, and is prone to mismatches on edge-case categories and unconventional products. The bigger hidden risk: if your existing taxonomy already has redundancy and errors, no matter how much the AI "hallucinates," wrong stays wrong — this method optimizes the workflow, not the source.

Impact on regular people

  • For enterprise IT: Small and medium businesses with product catalogs, content libraries, or document repositories can borrow this approach for low-cost tag cleanup, without immediately deploying dedicated classification systems.
  • For working professionals: People in content operations and e-commerce operations should remember this "AI drafts, vector verifies" human-machine division-of-labor template.
  • For consumers: E-commerce search and recommendations may get more accurate, but you might occasionally see "you searched for a coffee table, but it's recommending a side table" mismatches — glance at the category tree when shopping.