Context selection obeys the ordering that governs tools: permitted first, relevant second. A document your policy excludes cannot be reintroduced by a similarity score.
Retrieval runs over the set a request is allowed to see, not over everything indexed. Filtering after ranking is the common shortcut and the reason semantic systems leak — by the time the filter runs, the excluded content has already influenced what was selected.
Context is selected under an explicit token budget with mandatory inclusions honoured first. Retrieval that quietly triples prompt size converts a quality feature into a cost problem.
Embedding is provider-neutral: a managed model, your own key, or a self-hosted endpoint speaking the same shape are all valid, because some data classes must not leave the estate at all.
The embedding layer and vector storage exist and are tested; retrieval-backed knowledge bases are not yet a product surface. Nothing here is enabled during beta.