Quality Monitoring - Lumina Documentation

Detect quality degradation with hybrid scoring.

Scoring Methods

Exact Match:

Hash-based comparison
Instant results
Best for structured output

Semantic Similarity:

AI-powered scoring
~100ms latency
Best for natural language

Automatic Baselines

Lumina automatically:

Calculates baselines from history
Compares new responses
Alerts on degradation

Quality Alerts

Triggered when similarity drops:

Baseline: 0.92
Current: 0.65 (-29%)
Threshold: -20%
→ Alert sent

Custom Metrics

Add your own quality scores:

await lumina.traceLLM(
  () => llm.generate(prompt),
  {
    metadata: {
      factual_accuracy: calculateAccuracy(response),
      relevance: calculateRelevance(response),
    },
  }
);