med-routing — evaluate on your data

1 — upload your eval data

CSV with columns question (required), ground_truth, subject, difficulty, qid (optional). If ground_truth is present we skip generation; otherwise step 2 generates it with a frontier model.

2 — generate ground truth demo mode — cached

3 — evaluate cascade routers on this data

Sweeps thresholds for each selected router. Each cascade decision is scored against your ground truth by an LLM-as-judge (gpt-4o-mini). Output is a cost-vs-accuracy frontier — you pick your operating point.

4 — eval report

Eval observations are pushed to Prometheus at the recommended threshold for each router, so the dashboard reflects this run. Re-run eval to refresh.

all configurations