Key Performance Indicators (KPIs) and their implementation are crucial for the continuous improvement of manufacturing systems (MSs). In this article, we develop a KPI framework as a means to implement KPI computations in a convenient way. The idea is to go beyond the simple, standard definition of the KPIs by linking them explicitly with their instrumentation, that it to say the data required to calculate them. To do so, we present a set of metamodels: a KPI metamodel, a KPI-capture metamodel and a KPI-instrumentation metamodel. These are defined independently from any MS implementation, but are generic enough to be implemented in various Manufacturing Execution Systems. The use of the framework is illustrated on the Lampex learning factory from Strasbourg University.