HyDE: Hypothetical Document Embeddings

~15 min read

Questions aren't semantically similar to their answers, which hurts naive retrieval. HyDE fixes this by having the LLM generate a hypothetical answer first, then embedding THAT to search — even though the hypothetical answer may contain hallucinated details.

HyDE: Hypothetical Document Embeddings is a Pro topic

Sign in, then upgrade to Pro or Power to unlock this topic and the full AI Engineering curriculum.

Key points

  • Root problem: questions and their answers aren't semantically similar, so directly embedding the question retrieves irrelevant context
  • HyDE's fix: generate a hypothetical answer first, embed THAT instead of the question, then retrieve using the hypothetical's embedding
  • The hypothetical answer can and often does contain hallucinated details — this doesn't badly hurt retrieval
  • The contriever embedding model acts like a near-lossless compressor, filtering hallucinated specifics while preserving semantic shape
  • Trade-off: HyDE improves retrieval quality but adds latency and cost from the extra LLM call needed before retrieval even starts