This is interesting. Are you taking in public data sets in order to have your data accuracy or is the data set your using the proprietary advantaged here?
Agree with your notion of context bloat, have you though about the concept of context bundles and GraphDBs? This helped me a ton with limiting context and snapshots based on tooling.