RAG evaluation workbench

The problem

Useful experiments need a falsifiable question. This project begins by defining the behavior it is trying to understand, the constraints it holds constant, and the evidence that would change the approach.

The design

The implementation separates domain decisions from provider-specific infrastructure. Instrumentation is part of the first working path so latency, quality, cost, and failure are visible rather than reconstructed later.

What I learned

The valuable output is not only the code. It is the narrower understanding of when the approach works, where it breaks, and which complexity is justified in a production system.