Maybe align the architecture diagram in the readme? Having the boxes not align dure to the llm not being able to count correctly looks kind of bad.

I think however the idea with the knowledge graph is nice.