Exa Observability

Exa Observability

Output

• Bounded redacted metrics/traces for request outcome, latency, rate headroom, policy decision, source/citation coverage, and fallback behavior.
• An owned alert/runbook path that excludes raw queries, result content, identifiers, and credentials.

Examples

Emit aggregate metrics by environment, endpoint, status class, and policy outcome, then trigger a sanitized staging failure to verify alert routing. Record correlation ID and remediation time only; do not put customer queries, result text, or API keys into logs, dashboards, or tickets.

Overview

Monitor Exa search API performance, result quality, and cost efficiency. Key metrics: search latency by type (neural ~500-2000ms, keyword ~200-500ms), result count per query, cache hit rates, error rates by status code, and daily search volume for budget tracking.

Prerequisites

• Exa API integration in production
• Metrics backend (Prometheus, Datadog, or OpenTelemetry)
• Alerting system (PagerDuty, Slack, or equivalent)

Instructions

Step 1: Instrument the Exa Client

import Exa from "exa-js";

const exa = new Exa(process.env.EXA_API_KEY);

// Generic metrics emitter (replace with your metrics library)
function emitMetric(name: string, value: number, tags: Record<string, string>) {
  // Prometheus: histogram/counter.observe(value, tags)
  // Datadog: dogstatsd.histogram(name, value, tags)
  // OpenTelemetry: meter.createHistogram(name).record(value, tags)
  console.log(`[metric] ${name}=${value}`, tags);
}

async function trackedSearch(query: string, options: any = {}) {
  const start = performance.now();
  const type = options.type || "auto";
  const hasContents = options.text || options.highlights || options.summary;

  try {
    const method = hasContents ? "searchAndContents" : "search";
    const results = hasContents
      ? await exa.searchAndContents(query, options)
      : await exa.search(query, options);

    const duration = performance.now() - start;

    emitMetric("exa.search.duration_ms", duration, { type, method });
    emitMetric("exa.search.result_count", results.results.length, { type });
    emitMetric("exa.search.success", 1, { type });

    return results;
  } catch (err: any) {
    const duration = performance.now() - start;
    const status = String(err.status || "unknown");

    emitMetric("exa.search.duration_ms", duration, { type, status });
    emitMetric("exa.search.error", 1, { type, status });

    throw err;
  }
}

Step 2: Track Result Quality

// Measure whether search results are actually used downstream
function trackResultUsage(
  searchId: string,
  resultIndex: number,
  action: "clicked" | "used_in_context" | "discarded"
) {
  emitMetric("exa.result.usage", 1, {
    action,
    position: String(resultIndex),
  });
  // Results at position 0-2 should have high usage
  // If top results are discarded, query needs tuning
}

// Track content extraction value
function trackContentValue(result: any) {
  if (result.text) {
    emitMetric("exa.content.text_length", result.text.length, {});
  }
  if (result.highlights) {
    emitMetric("exa.content.highlight_count", result.highlights.length, {});
  }
}

Step 3: Cache Monitoring

class MonitoredCache {
  private hits = 0;
  private misses = 0;
  private cache: Map<string, { data: any; expiry: number }> = new Map();

  async search(exa: Exa, query: string, opts: any) {
    const key = `${query}:${opts.type}:${opts.numResults}`;
    const cached = this.cache.get(key);

    if (cached && cached.expiry > Date.now()) {
      this.hits++;
      emitMetric("exa.cache.hit", 1, {});
      return cached.data;
    }

    this.misses++;
    emitMetric("exa.cache.miss", 1, {});

    const results = await exa.searchAndContents(query, opts);
    this.cache.set(key, { data: results, expiry: Date.now() + 3600 * 1000 });
    return results;
  }

  getStats() {
    const total = this.hits + this.misses;
    return {
      hits: this.hits,
      misses: this.misses,
      hitRate: total > 0 ? `${((this.hits / total) * 100).toFixed(1)}%` : "N/A",
    };
  }
}

Step 4: Prometheus Alert Rules

groups:
  - name: exa_alerts
    rules:
      - alert: ExaHighLatency
        expr: histogram_quantile(0.95, rate(exa_search_duration_ms_bucket[5m])) > 3000
        for: 5m
        annotations:
          summary: "Exa search P95 latency exceeds 3 seconds"

      - alert: ExaHighErrorRate
        expr: rate(exa_search_error[5m]) / rate(exa_search_success[5m]) > 0.05
        for: 5m
        annotations:
          summary: "Exa API error rate exceeds 5%"

      - alert: ExaEmptyResults
        expr: rate(exa_search_result_count{result_count="0"}[15m]) > 0.2
        for: 10m
        annotations:
          summary: "Over 20% of Exa searches returning empty results"

      - alert: ExaCacheHitRateLow
        expr: rate(exa_cache_hit[5m]) / (rate(exa_cache_hit[5m]) + rate(exa_cache_miss[5m])) < 0.3
        for: 15m
        annotations:
          summary: "Exa cache hit rate below 30% — check query patterns"

Step 5: Health Check Endpoint

app.get("/health/exa", async (_req, res) => {
  const start = performance.now();
  try {
    const result = await exa.search("health check", { numResults: 1 });
    const latencyMs = Math.round(performance.now() - start);
    res.json({
      status: "healthy",
      latencyMs,
      resultCount: result.results.length,
    });
  } catch (err: any) {
    res.status(503).json({
      status: "unhealthy",
      error: err.message,
      latencyMs: Math.round(performance.now() - start),
    });
  }
});

Dashboard Panels

Error Handling

Resources

• Exa API Documentation
• Exa Rate Limits
• Prometheus Alerting Rules

Next Steps

For incident response, see exa-incident-runbook. For cost optimization, see exa-cost-tuning.