Retrieval

Retrieval is the only process that touches a vector store. Nothing else names a collection, a table, or a filter syntax.

The scope filter

Every read and every delete carries a filter built in one place, from the request's library and generations. It is never assembled at a call site.

Isolation that depends on each caller remembering to add a condition holds right up until the first caller forgets.

Scope is required and has no default. A request that could omit its library is refused before any store is reached, because on the day someone forgot, an empty filter would return another user's documents.

The principal is audit context

Requests carry the acting principal, but Retrieval does not use it to decide access. Whether that caller may read that library was decided by Control before the request was made; re-deciding it here, without Control's data, would be guessing.

Generations

A generation is one build of one document version's index.

Searches read only the generation Control marks active, so a rebuild in progress is invisible rather than partially answering questions. Activation is a single row update, so every answer changes at the same instant.

A build that is short of the passages it should hold never activates. Activating it would drop the missing passages from every future answer, with nothing to show that anything was lost — so the index stage's whole job is to count, compare, and refuse.

Cleanup

Retiring a superseded build and erasing a deleted document are the same operation with one parameter different, so they share a code path rather than two that must stay in agreement about what "gone" means.

Cleanup addresses a version rather than a generation, because the builds a version has been through are not recorded anywhere and the point of cleanup is to leave nothing behind.

Superseded generations are retired after activation, never before. Until Control has switched, the older generation is what searches are reading.

No Primer vector adapter

Haystack's DocumentStore protocol is already that abstraction. Wrapping it would add a layer whose only purpose would be to become a second place for isolation bugs to live.

The conformance suite

pgvector and Qdrant are both supported, so both run the same tests: write, filtered search, cross-library and cross-owner isolation, duplicate indexing, pending-generation invisibility, generation swap, delete, repeated delete, dimension mismatch, and citation round-trip.

The isolation cases put the same words in two libraries, so nothing but the filter separates them. A backend that ignored it fails there rather than in production.

That suite has already earned its place: it caught the pgvector integration naming its indexes with fixed defaults, which makes two Primer tables in one schema collide on creation.