OpenEvidence Observability

Overview

OpenEvidence delivers clinical evidence queries where response accuracy and freshness have direct patient safety implications. Monitor query response times to ensure clinicians get timely answers, track evidence freshness to catch stale citations, and audit every query for compliance. Observability must also verify citation accuracy and maintain complete audit logs for regulatory requirements (HIPAA, clinical decision support standards).

Key Metrics

Metric	Type	Target	Alert Threshold
Query response time p95	Histogram	< 3s	> 8s
Evidence freshness	Gauge	< 7 days median	> 30 days
Citation accuracy rate	Gauge	> 95%	< 90%
API error rate	Gauge	< 0.5%	> 2%
Audit log completeness	Gauge	100%	< 99.9%
Daily query volume	Counter	Within quota	> 90% quota

Instrumentation

async function trackClinicalQuery(queryType: string, fn: () => Promise<any>) {
  const start = Date.now();
  const traceId = crypto.randomUUID();
  try {
    const result = await fn();
    metrics.histogram('openevidence.query.latency', Date.now() - start, { queryType });
    metrics.increment('openevidence.query.total', { queryType });
    auditLog.record({ traceId, queryType, status: 'ok', latency: Date.now() - start });
    return result;
  } catch (err) {
    metrics.increment('openevidence.query.errors', { queryType, error: err.code });
    auditLog.record({ traceId, queryType, status: 'error', error: err.message });
    throw err;
  }
}

Health Check Dashboard

async function openEvidenceHealth(): Promise<Record<string, string>> {
  const latencyP95 = await metrics.query('openevidence.query.latency', 'p95', '5m');
  const errorRate = await metrics.query('openevidence.query.error_rate', 'avg', '5m');
  const freshness = await openEvAdmin.getMedianEvidenceAge();
  return {
    query_latency: latencyP95 < 3000 ? 'healthy' : 'slow',
    error_rate: errorRate < 0.005 ? 'healthy' : 'degraded',
    evidence_freshness: freshness < 7 ? 'healthy' : 'stale',
  };
}

Alerting Rules

const alerts = [
  { metric: 'openevidence.query.latency_p95', condition: '> 8s', window: '10m', severity: 'warning' },
  { metric: 'openevidence.query.error_rate', condition: '> 0.02', window: '5m', severity: 'critical' },
  { metric: 'openevidence.evidence.median_age_days', condition: '> 30', window: '1d', severity: 'warning' },
  { metric: 'openevidence.audit.completeness', condition: '< 0.999', window: '1h', severity: 'critical' },
];

Structured Logging

function logClinicalEvent(event: string, data: Record<string, any>) {
  console.log(JSON.stringify({
    service: 'openevidence', event,
    query_type: data.queryType, duration_ms: data.latency,
    citation_count: data.citations, evidence_age_days: data.evidenceAge,
    // HIPAA: never log patient identifiers or query text
    trace_id: data.traceId, audit_seq: data.auditSeq,
    timestamp: new Date().toISOString(),
  }));
}

Error Handling

Signal	Meaning	Action
Query timeout > 8s	Evidence index overloaded	Check index health, scale read replicas
Citation accuracy drop	Stale or retracted sources	Trigger evidence refresh pipeline
Audit log gap	Logging pipeline failure	Critical — investigate immediately for compliance
429 rate limit	Quota approaching limit	Throttle non-critical queries, request increase
Evidence age > 30 days	Refresh pipeline stalled	Check ingestion jobs, verify source feeds

Resources

• OpenEvidence

Next Steps

See openevidence-incident-runbook.