We built one of the largest RAG set-ups that exist toady with Usul.ai (6B tokens). We started by using langchain and llamaindex, they were able to get us to a prototype in a couple of days, but took 3 months of taking pieces apart and optimizing them to make it perform well at such large scale.
We put all of these learning into an MIT licensed open-source project — Agentset. Our goal to let people get production quality RAG w/o having to understand or optimize the underlying pieces. It supports 22 file formats, agentic search, deep research, citations, and a UI out of the box.
Happy to answer any questions about our journey or the product.