Vector embeddings solve LLM tagging at scale without fixed vocabularies
Doug Turnbull proposes a pragmatic workaround for LLM-based content tagging at scale. Rather than constraining models to a fixed vocabulary (which becomes unwieldy with thousands of tags), he suggests letting models generate novel tags freely, then anchoring those outputs to existing categories via vector similarity. This inverts the typical classification pipeline: instead of choosing from a predefined set, the model hallucinates plausible labels and retrieval finds the nearest canonical match. The technique sidesteps both prompt-injection risks and token-budget constraints, making it valuable for anyone managing large taxonomies or legacy content needing retroactive organization.
Modelwire context
ExplainerTurnbull's insight is that the constraint itself is the problem. Rather than fighting LLMs' tendency to generate novel outputs, he weaponizes it: let the model produce whatever tags feel right, then use vector similarity to snap them to your canonical taxonomy post-hoc. This is less about a new algorithm and more about accepting model behavior instead of fighting it.
This is largely disconnected from recent activity in the space, which has focused on either making models more obedient (constitutional AI, RLHF) or measuring their failures (jailbreaks, adversarial robustness). Turnbull's framing belongs to a quieter conversation about pragmatic deployment: how do you actually ship systems that work with LLM quirks rather than against them. It sits alongside operational work on prompt caching and token budgeting rather than capability benchmarks.
If teams managing large e-commerce or content taxonomies (Shopify, Notion, Atlassian) ship this pattern in their LLM-powered tagging features within the next 12 months, it signals this has moved from clever workaround to standard practice. If it remains a one-off technique in Doug's blog, it's a useful idea without adoption momentum.
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsDoug Turnbull · Simon Willison · LLM
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. Simon Willison originally reported this story as “Don't classify. Hallucinate!”. The full content lives on simonwillison.net. If you’re a publisher and want a different summarization policy for your work, see our takedown page.