HyDE generates a short hypothetical answer document for a query, embeds that text, and uses it to retrieve real passages. It is useful when users ask vaguely and a raw query embedding would not match how documents are written.
Whilst HyDE improves recall on fuzzy questions, in practice the hypothesis can invent entities the index then “confirms.” For example, a made-up product name retrieves lookalike pages and the final answer treats them as fact. We often recommend HyDE only for candidate retrieval — then re-rank and answer strictly from verified sources.
Query embeddings and document embeddings live in the same space but differ in style. HyDE narrows that gap by generating a short document-shaped hypothesis, embeddingthat, and searching. It boosts recall on vague or abstract questions and can amplify invented entities if you treat hypothesis hits as truth. Production pattern: HyDE for candidates → hybrid/RRF → re-rank → generate only from real sources. Log the hypothesis text in evals so you can see when it helps versus harms.