Article · 2026-08-14 · 4 moments

Don't classify. Hallucinate!

Don't classify. Hallucinate! I still have quite a bit of older content on my blog that I never got round to tagging. My blog has 1,856 tags - likely too many to feed to an LLM in one go and say "which of these tags match the following content". Doug Turnbull has a neat solution. Tell the model to output tags without any details of the existing vocabulary, then use vector embeddings against the existing corpus to find the concrete tags that are closest to the ones the model imagined might fit! ✦ AI generated

Read original ↗

GUEST the author GUEST the author (quoting Doug Turnbull's example prompt)HOST Simon Willison

Context

A tag vocabulary of 1,856 entries is too large to feed to an LLM in one shot as a flat multiple-choice 'which of these tags match' prompt.

Untagged older blog content sits against an 1,856-tag vocabulary that is impractical to hand to an LLM whole and ask it which tags fit.

transcript

the author: I still have quite a bit of older content on my blog that I never got round to tagging. My blog has 1,856 tags - likely too many to feed to an LLM in one go and say "which of these tags match the following content".

gives example · 1

Claim

Including an example of the shape of your tags helps the model produce a more useful hallucinated guess.

The prompt should include an example of the tag format — like hierarchical product classifications such as 'Furniture / Living Room Furniture / Coffee Tables' — so the model's imagined tags look right.

transcript

the author: His example prompt suggests including an example of the shape of your tags to help the model make a more useful guess.

Mechanism◆

Instead of classifying against a fixed tag vocabulary, let the model freely hallucinate candidate tags, then resolve them to concrete existing tags via vector embeddings over the corpus.

Doug Turnbull's approach: have the model emit imagined tags unconstrained by the existing vocabulary, then use vector embeddings against the corpus to find the concrete tags closest to what it invented.

transcript

the author: Doug Turnbull has a neat solution. Tell the model to output tags without any details of the existing vocabulary, then use vector embeddings against the existing corpus to find the concrete tags that are closest to the ones the model imagined might fit!

explains mechanism · 2gives example · 1

Example

The recommended prompt instructs the model to create novel, never-before-seen product classifications that best fit a search query.

An illustrative prompt: tell the model to devise novel furniture/home-goods classifications fitting a query, given example classification shapes, then supply the query (e.g. 'brown coffee table').

transcript

the author (quoting Doug Turnbull's example prompt): Your task is to create novel, never seen before, furniture, home goods, or hardware classification that best fit a search query. ... Here's the query to generate classifications for: brown coffee table

Highlight slides

Freely hallucinate tags, resolve via embeddings✦ from: Instead of classifying against a fixed tag vocabulary, let the model freely hallucinate candidate tags, then resolve them to concrete existing tags via vector embeddings over the corpus.

Related episodes

Video · Syntax

I Built an LLM From Scratch

→

Video · Lenny's Podcast

The rise of taste, human authenticity and judgment in an AI world | Adam Mosseri (Head of IG)

→

Video · The TWIML AI Podcast

Teaching Computers to Smell | Alex Wiltschko

→