- https://arxiv.org/pdf/2212.10496
- Source: arxiv
TL;DR
HyDE (Hypothetical Document Embeddings) achieves zero-shot dense retrieval competitive with supervised systems by having an LLM generate a hypothetical answer document and embedding it with Contriever for nearest-neighbor search.
Summary
HyDE solves zero-shot dense retrieval by inverting the usual query-to-document matching direction. An instruction-following LLM generates a hypothetical answer document for the query — plausible in style and structure even if factually imprecise. A contrastive encoder (Contriever) then embeds that hypothesis into the document space, where nearest-neighbor search retrieves real passages. The hypothesis acts as a soft relevance signal that aligns query intent with document vocabulary, matching supervised dense retrieval on BEIR without any labeled data.
Key contributions
- Proposes HyDE, a zero-shot retrieval method that replaces query encoding with encoding of a hypothetically generated answer document.
- Demonstrates that factual correctness of the hypothesis is irrelevant — only its embedding proximity to real documents matters.
- Matches supervised dense retrieval performance on BEIR without any labeled relevance data.
- Shows that larger instruction-following LLMs produce better hypotheses and improve recall, linking generation quality to retrieval quality.
When to cite
- When implementing zero-shot dense retrieval without labeled training data for a new domain.
- When arguing that LLMs are useful as retrieval preprocessors even when they hallucinate, because factual accuracy is decoupled from embedding utility.
- When evaluating HyDE as a strong baseline before investing in fine-tuned retrievers or hard-negative training.
- When discussing query expansion techniques that require no thesaurus or domain-specific vocabulary resources.