RAG demos are easy to make impressive. Reliable RAG is harder to measure.
For Cortagent, public claims need proof. That means ground truth, reproducible tests, documented failures, and environment details for latency.
Accuracy proof needs
- a fixed task set,
- expected answers,
- source material,
- scoring rules,
- failure categories,
- and rerunnable evaluation.
Latency proof needs
- environment description,
- model and retrieval configuration,
- repeated runs,
- variance reporting,
- and separation between cold and warm paths.
No benchmark laundering
A benchmark without methodology is not evidence. A single good run is not a result.
Why publish this position
Agentic RAG is complex enough that vague metrics are misleading. The measurement system has to show not only whether the answer was accepted, but why retrieval and evidence selection supported it.