Wait what, the conversion is so lossy you use an LLM to analyse it? Or you get the claude analysis into JSON and put that into the system too? I'm not seeing what's causing the 2-3x efficiency bump.