How would you support a query like ‘what happened right before the manager approved the budget’ against a video index?

Since standard embedding techniques often lose temporal ordering, filtering by a timestamp column after retrieval is insufficient because it may return topically relevant but temporally incorrect results. By applying a rotation to each event’s embedding proportional to its timestamp, you encode the temporal position directly into the vector space. This ensures that the similarity search itself accounts for the relative elapsed time between events.