What's your top 3 techniques that helped you save tokens / better accuracy from models?
Focusing on curating the context, keeping it under 150k.
Making tools to trim git / grep / ls output to my likingindexing my large code base and making the basic tools show short index options + hints. Custom tools was a massive red herring and waste of time, models hate using them and never use them properly
live filter on my codebase to sandbox models. i.e. exclude file types, files older than N days. etc
cosmicnag Was using blackhole earlier, trying billion context now - when it works, its like blackhole on autopilot 4d ago
137_1 I would say RTK. 5d ago
wagwan_g112 I’m fairly sure RTK was debunked for actually increasing token usage due to the model needing to repeat failed or truncated tool calls. 5d ago 2 likes