Under maintenance: the live demo backend is being upgraded, so queries may fail temporarily. The Tech Wiki and Metrics pages are unaffected.

Back to research

Quality Metrics

Every eval run pushes the agent through a golden dataset of research queries, scores the answers with an independent LLM judge (Gemini, a different model family than the agent's Llama), and checks whether the agent's self-critique agrees with the external judge.

Run: 2026-07-06Judge: gemini-2.5-flashAgent LLM: groq1/2 queries scoredbundled snapshot
Faithfulness
0.36/ 1.00

Are the answer's claims grounded in retrieved context? (ragas, LLM judge)

Context Precision
1.00/ 1.00

How much of the retrieved context is actually relevant? (ragas)

Fact Recall
1.00/ 1.00

Fraction of expected golden facts covered by the answer

Answer Quality
0.50/ 1.00

Rubric-scored relevance, completeness, specificity, coherence

Critic Calibration

Pearson correlation between the agent critic's own quality score and the external judge. Higher means the agent knows when its own answers are good.

Avg Self-Score

0.80

Agent critic's average confidence

Avg Iterations

2.0

Self-reflection loops per query

Avg Latency

104s

End-to-end per research query

Per-Query Results

What is Microsoft's corporate mission?

Error: judge: Gemini free-tier daily quota exhausted (429 RESOURCE_EXHAUSTED)