Faithfulness and relevancy, scored against ground truth.
Every response is graded on whether its claims are actually supported by retrieved context, and whether it answers the question that was asked. These are live telemetry from evaluation runs, not projected targets.
Faithfulness
Claims supported by context
Answer relevancy
Response addresses the query
3
queries scored
Methodology3-query pilot evaluation. Groq's API does not support n>1, so each metric reflects a single sampled generation per query rather than an averaged multi-sample estimate. Read these as directional, not statistically robust.
Query level breakdown
MeasuredWhat are NVIDIA's primary strategies for mitigating supply chain constraints?
If NVIDIA's reported data center revenue drops by 10% from $47 billion, what is the new figure?
What does the 10-K say about NVIDIA's secret plans to acquire AMD?