Sources
'Don't Classify. Hallucinate!' — Doug Turnbull's Trick for Tagging Against a 1,856-Tag Vocabulary Too Big to Fit in a Prompt
Simon Willison highlighted Doug Turnbull's technique on August 14, 2026 for a problem every RAG builder hits: his blog has 1,856 tags, far too many to hand an LLM as a constrained vocabulary. Instead of classifying, you let the model freely invent tags with no vocabulary at all, then embed its inventions and nearest-neighbour them back onto the real corpus. Willison notes the prompt still shapes the output by showing structural examples like 'Furniture / Living Room Furniture / Coffee Tables' — the model needs the shape, just not the list.
↳ Follow the thread