Do you think it makes more sense to apply semantics when data is generated, or further down the data/analytics pipeline?

I think the approach varies depending on the nature of the data, such as whether it is structured or unstructured, and the quantity of data. In the healthcare domain where I work, I think performing semantic redefinition is most effective later in the pipeline. I would be concerned about performance if it were done upfront, especially given the terabytes of data we receive daily, much of which is non-discrete and unstructured. I also appreciate the flexibility to modify the semantic model later in the pipeline without the need to regenerate data.