Query Expansion with HyDE

HyDE operates on a grounding-by-generation principle. Instead of directly embedding a short, ambiguous query, a language model first generates a hypothetical ideal document that answers the query. The dense embedding of this synthetic, detailed text is then used to perform a vector similarity search against the real document corpus. This process effectively re-anchors the query in the embedding space, moving it closer to the region where actual relevant documents reside, without requiring any task-specific training data.

Traditional query expansion methods, such as Pseudo-Relevance Feedback (PRF) or synonym expansion, add discrete terms to the original query. HyDE fundamentally differs by generating a complete, coherent passage. Key distinctions:

• PRF assumes top-k initial results are relevant, risking query drift from bad initial hits.
• Synonym expansion only adds words with similar meanings, failing to capture complex contextual relationships.
• HyDE leverages the language model's parametric knowledge to synthesize a document that captures the query's intent in the same verbose, factual style as the target corpus, even if the generated text contains fabricated details.

The instruction-tuned language model acts as a query-to-document translator. It is prompted with a simple instruction like, 'Write a passage that answers the question.' The model's output is not required to be factually correct; its purpose is to capture the linguistic pattern and topical density of a relevant document. This synthetic document serves as a dense, high-dimensional proxy for the user's information need, effectively performing a form of unsupervised domain adaptation for the retrieval system.

A core advantage of HyDE is its unsupervised alignment with the target corpus. The generative model produces text in a verbose, explanatory style that naturally mirrors the statistical patterns of the indexed documents. This contrasts sharply with embedding a terse query directly, which often maps to a sparse, low-information region of the vector space. By generating a document-like embedding, HyDE performs a zero-shot transformation that aligns the query representation with the distribution of the document embeddings, significantly improving dense retrieval recall.

The behavior of HyDE can be steered by modifying the instruction prompt given to the generative model. For a legal corpus, the prompt might be: 'Write a legal memorandum answering the question.' For a medical corpus: 'Write a section of a clinical guideline that addresses the topic.' This simple prompt engineering allows the same core architecture to adapt its generated document style to different domains, controlling the linguistic register and topical specificity of the hypothetical document to better match the target index.

While HyDE is primarily a retrieval-stage technique, its generated document can also be used for re-ranking. The synthetic passage can be fed alongside a candidate document into a cross-encoder model to compute a fine-grained relevance score. This provides a powerful signal because the cross-encoder can directly compare the generated 'ideal' answer with the actual retrieved document, assessing factual overlap and logical consistency beyond simple vector similarity.