What the model is allowed to draw on, and how it is grounded. Retrieval over a private semantic index, an app's own data, personal context — answers anchored to real, attributed sources instead of the model's guesses.
A model left to itself answers from the distribution it was trained on — fluent, plausible, and unattributable. Context Intelligence is the discipline of giving it something better to stand on: retrieval over a private semantic index, an app's own entities, the user's real data, surfaced at the moment of the question. Apple's framing is concrete — entity schemas feed the Spotlight semantic index, and a RAG tool backed by Core Spotlight lets a model ground its answers in local, on-device content.
The discipline is not “add a vector database.” It is deciding what the model may draw on, keeping that retrieval on-device and attributed, and ensuring every grounded claim can be traced to the source it came from. Context that cannot be attributed is just a more confident guess.
What the model knows should be retrieved, grounded, and attributed — never assumed.
Observe gathers the external field — signal arriving from the world. Context grounds the model in what is already known: the user's data, the app's entities, the private index. Different direction; the same demand for attribution.
The essay, the build, and the field notes arrive in dedicated sessions. This page states what is coming rather than pretending it is already here.
What the model may draw on, kept local and attributed. Grounding is the work.
Observe gathers the external field; Context grounds the model in what is already known. The two inputs to any honest answer.
Observational Intelligence →What the model may retrieve is a governance question: which sources, kept where, and attributed how. Grounding without governance is just a wider blast radius.
Governed Intelligence →Grounding belongs on the device — a private index the answer never has to leave home to consult.
Edge Intelligence →