Hypothetical Document Embeddings
Also called HyDE method
Search by first imagining the answer, then finding real documents that look like it.
Think of it like
Sketching what the missing puzzle piece should look like, then hunting for the piece that matches your sketch.
Example
For "why is my sourdough dense?", the model drafts a plausible answer, embeds that, and retrieves real baking docs similar to the draft.
How it actually works
Short queries embed poorly because they carry little signal. This technique asks an LLM to generate a hypothetical answer, embeds that richer text, and retrieves against it — the fake document is closer in vector space to the real ones than the bare query was. It boosts recall on tough questions, but if the model hallucinates the wrong kind of answer, it can steer retrieval astray.
For product teams
A clever recall boost for vague or terse queries; adds an LLM call and can misfire if the draft is off-base.
For engineers
Generate a pseudo-answer, embed it, retrieve by its vector; bridges the query–document distribution gap in dense retrieval.
Related
- HyDE — Better known by its acronym, HyDE.
- Dense Retrieval — Improves dense retrieval on hard queries.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome