Don't classify. Hallucinate!
Use LLM-hallucinated tags plus vector embeddings to match against large existing vocabularies
Doug Turnbull proposes skipping constrained classification prompts and instead letting an LLM freely invent tags, then using vector similarity to map those invented tags onto an existing corpus vocabulary. This sidesteps the problem of feeding thousands of candidate labels to a model in a single prompt. It's a practical retrieval-augmented classification trick useful for content tagging at scale.