ai · 1 min read
HyDE: Generating Hypothetical Documents to Improve Retrieval
HyDE asks the LLM to write a hypothetical answer, then embeds that. Counterintuitive - and remarkably effective. Here is when to use it.
BABhogar AI TeamProduct & Engineering
Hypothetical Document Embeddings (HyDE) asks the LLM to write what an answer might look like, then retrieves documents similar to that hypothetical answer. It sounds backward and works surprisingly well.
Why it matters
HyDE shines on exploratory and zero-shot queries where the user query is far from the language of the corpus. It is overkill for short factoid lookups and burns tokens.
How Bhogar AI approaches it
Bhogar AI offers HyDE as a per-query strategy invoked when the rewriter detects high query-corpus distance. You get the recall lift on hard queries without paying the token tax on easy ones.
- Per-query strategy selection
- Distance-based gating between vanilla and HyDE
- Compatible with hybrid retrieval
- Per-strategy telemetry in eval dashboard
- Hard budget cap per query
What you get
On hard exploratory queries HyDE lifts top-3 recall 15-25 points; on easy factoid queries the gating skips it and saves tokens.