Matching the person's current task context — an open document, an active channel — against likely-relevant content, to surface a suggestion unprompted.
Embedding the current context and matching it against the index continuously, with a relevance threshold gating when to interrupt.
Same embedding infrastructure as CAP-102