HyDE Hypothetical Embeddings

Hypothetical Document Embeddings (HyDE) is a Zero-Shot retrieval enhancement technique that bridges semantic distribution gaps between short queries and long documents; given user query $q$, HyDE prompts an offline LLM to synthesize a 'Hypothetical Document' $d_{\text{hypo}}$; even if factually ungrounded, its vocabulary distribution and syntactic domain align closely with true target documents; $d_{\text{hypo}}$ is then embedded into the vector store, converting asymmetric 'query-to-doc' search into symmetrical 'doc-to-doc' retrieval.

Short ambiguous queries, exploratory open-domain search, and cross-domain technical documentation lacking explicit keywords.

Short queries suffer severe Query-Doc Embedding Asymmetry, yielding weak cosine similarity against dense paragraph vectors; HyDE expands sparse query intents into full-length simulated documents, unlocking high-fidelity symmetric embedding alignment.

Why does HyDE successfully retrieve correct real documents even when the synthesized hypothetical document contains factual hallucinations?

Under what specific scenarios does HyDE completely fail (e.g., proprietary acronyms, private code repos, cold-start entities)?

How can small language models (1B-3B SLMs) mitigate the TTFT latency penalty introduced by HyDE's preliminary generation step?

How does Multi-HyDE (sampling 3-5 hypothetical docs and averaging embeddings) stabilize retrieval variance over single-doc HyDE?

What is the engineering heuristic for weighted fusion ($v_{\text{final}} = \alpha v_q + (1-\alpha) v_{\text{hypo}}$) of query and hypothetical vectors?