October 20, 2025 • 3 min readI've spent the last 8 months in the RAG trenches, I want to share what actually worked vs. wasted our time. We built RAG for Usul AI (9M pages) and an unnamed legal AI enterprise (4M pages). Langchain + Llamaindex We started out with youtube tutorials. First Langchain → Llamaindex. Got to a working prototype in a couple of days and were optimistic with the progress. We run tests on subset of the data (100 documents) and the results looked great. We spent the next few...