Short for Hypothetical Document Embeddings, HyDE is a retrieval technique from a 2022 arXiv paper: rather than embedding a terse user query directly, an LLM first writes a plausible (possibly wrong) hypothetical answer or document, and that hypothetical text is embedded and compared against real documents. Because the hypothetical answer is written in the same register and detail level as real matching content, its embedding often lands closer to the correct document than the original query's embedding would. Developer Doug Turnbull applied the same idea to classification instead of search — hallucinating a plausible category label, then resolving it to the closest real category via embedding similarity — letting a cheap model classify into a large taxonomy without ever seeing the taxonomy itself.