HyDE
Also called Hypothetical Document Embeddings
The catchy name for retrieving with an imagined answer instead of the raw question.
Think of it like
Guessing what the answer key says, then finding the textbook page that matches your guess.
Example
A research tool uses HyDE so that "effects of microplastics on soil" pulls papers matching a drafted summary, not just the keywords.
How it actually works
HyDE is shorthand for Hypothetical Document Embeddings: prompt an LLM for a plausible answer, embed it, and use that vector to retrieve. It reliably lifts zero-shot dense retrieval because generated prose sits nearer the target documents than a short query does. The trade is an extra generation step and sensitivity to the draft’s quality.
For product teams
Same idea as its longer name — worth trying when short queries retrieve poorly.
For engineers
LLM-drafted pseudo-doc embedded for retrieval; strong zero-shot gains, dependent on generation fidelity.
Related
- Hypothetical Document Embeddings — The full name and mechanism.
- Retrieval — A trick layered onto standard retrieval.
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