Comment by tdu01

We built something similar for our team at work. A human has to run it and review/skim for correctness, who quickly gets overwhelmed. The purpose eventually became an AI wiki for AI to use to spawn agents more efficiently (token wise), to pull in only modules it needs and fills in the blanks from the md files. Like a knowledge graph. It has become useful for production support, design, feature dev, domain specific learnings. We also found it makes a huge difference which model you use to encode and which model you use to decode.