I’ve been building a Company Brain (https://himanshusharma.org/notes/building-a-company-brain/) and so far it’s been going good. While I was reading about Google’s OKF, I started to wonder what the limits of my Company Brain are.
For my personal use case, it will never 1 million customer tickets so RAG seems an overkill. A simple LLM wiki seems simple and easy to maintain. Not as flashy as saying I use Mem0, Pinecone and Supabase for my second brain, but it’ll do the job.
But for my clients, who have millions of docs, tickets, help articles, SOPs and reports, the Markdown files won’t work at that scale. At that scale, RAG is a no brainer, even with frontier agents getting more capable and handling context compaction better.
And do we even need frontier models for the Company Brain? I’ve been using GPT-6 Luna High and Deepseek Flash v4.1 for a while, and they seem good enough. I should spend some more time confirming my hunch that neither my clients nor I need SOTA models for the Company Brain.