Retrieval-augmented generation looks easy in a notebook and breaks in production. Here's how we make it citation-grade.
Retrieval-augmented generation (RAG) is the fastest way to ground an LLM in your own data — but the gap between a weekend demo and a reliable production system is enormous.
At NOYT India we build these systems in production — with evaluation, cost control, and observability baked in from day one. Below we share the patterns that have held up under real traffic, the trade-offs we weigh, and the mistakes worth avoiding.
If you're planning something similar, the short version is: start from the bottleneck, instrument everything, and ship in small, reversible steps.