Building a Go backend for orchestrating AI agents (multi-tenant, each agent has its own persona/tools/LLM). Now I'm stuck on how knowledge bases should work and I keep going back and forth between "make it flexible" and "just ship something simple." Here's where I landed, architecture-wise: Source = wherever the data lives. S3 bucket of PDFs, a website you crawl, a Notion workspace, whatever. Normalizer = takes whatever comes out of the source and turns it into some...