After building enterprise RAG from scratch, sharing what I learned the hard way. Some techniques I expected to work didn't, others I dismissed turned out crucial. Covers late chunking, hierarchical search, why reranking disappointed me, and the gap between academic papers and messy production data. Still figuring things out, but these patterns seemed to matter most.