Notes

Catching embedding drift before users do

Retraining an encoder silently changes the geometry of the index. Tracking the distribution of nearest-neighbour distances for a fixed probe set, plus overlap@k between old and new models, turned out to be the most useful early warning.