Really solid write-up — it’s rare to see someone break down the real tradeoffs of scaling RAG beyond the toy examples. The bit about reranking and chunking actually saving more than fancy LLM tricks hits home to me.