Evaluation dashboard
Compare a candidate against a baseline across quality, latency and cost.
baseline
- Faithfulness
- 0.81
- Recall@8
- 0.93
- p95 latency
- 1.9 s
- Cost / query
- $0.018
candidate
- Faithfulness
- 0.88
- Recall@8
- 0.91
- p95 latency
- 2.4 s
- Cost / query
- $0.023
Faithfulness
+7 pts
Recall
−2 pts
Latency
+0.5 s
The candidate is more faithful but slower, more expensive and slightly worse at recall. Ship only if the faithfulness gain matters more than the tail latency and recall loss.
Related: Why evaluation