RAG Evaluation System
An end-to-end retrieval-augmented generation pipeline — load, chunk, embed, store, retrieve, generate, serve — built around a first-class evaluation harness, so every configuration choice is measured rather than assumed.
The differentiator is the method. A 3×2 factorial experiment over chunking strategy and embedding model, with scoring rules pre-registered before any result was computed. McNemar's exact test on paired outcomes, bootstrap confidence intervals at 10,000 resamples, and a 25-point sensitivity sweep to confirm the conclusion survives rejecting its own weights. The README states which gains are statistically demonstrated and which are not: the winning configuration is not separable from its runner-up, and it says so.
