12 Things Nobody Tells You About Building a Production RAG System

I spent several months building Ask360, a multi-tenant RAG platform that lets anyone upload documents and deploy an AI assistant with source citations and confidence badges. The UI and CRUD took a few weeks. The search quality, failover, and production hardening? That's where the real engineering happened. Here are the problems I hit that tutorials don't cover, and the solutions that actually work in production. 1. You need two databases, and that's actually fine Every RAG tutorial shows one database. In production, I use two: MariaDB for user accounts, projects, documents, billing, and usage tracking. Everything the admin panel touches. PostgreSQL + pgvector for vector embeddings, semantic cache, and chunk storage. Everything the search engine touches. Why not one? MariaDB is excellent for the CRUD-heavy admin layer. Yii2's ActiveRecord, migrations, fixtures, and auth all work out of the box. But MariaDB has no vector similarity search. PostgreSQL with pgvector gives me HNSW indexes f