Private Brains: Retrieval-Grounded Enterprise AI Without Data Egress

An enterprise assistant is only as useful as its access to the organisation's own knowledge, yet the common way of providing that access, sending documents and their embeddings to a cloud service for indexing and retrieval, reintroduces exactly the data-egress problem that regulated organisations must avoid. This paper describes a retrieval architecture in which an organisation's knowledge is held in private knowledge bases, which we call brains, on hardware the organisation owns, and is made available to an offline assistant without any data leaving the building. Ingestion, indexing, routing and retrieval all run locally. A routing layer selects the brains relevant to a query and retrieves only the material needed, so that a single prompt's working context stays within the served model's context window while large corpora remain usable. We have exercised retrieval over a 41 GB reference corpus on a single workstation, which indicates that corpus size is bounded by local storage rather than by an external service. Retrieved sources are recorded in the system's attestation record, so that it can be shown after the fact which knowledge informed a given answer. We contrast the approach with cloud retrieval-augmented generation and discuss its operational and regulatory implications.