Why it matters

HyDE improves retrieval quality. Understanding shapes technique.

Advertisement

The architecture

LLM: generate hypothetical answer for query.

Embed hypothetical doc.

Retrieve nearest actual docs.

HyDE flowQueryuser inputLLM hypotheticalgenerated answerEmbed + retrieveagainst real docsGao et al. 2022; helps when query embedding style differs from doc style
HyDE math.
Advertisement

How it works end to end

Small LLM sufficient for hypothetical.

Multiple hypotheticals for diversity.

Combine with hybrid + rerank.