mistral-observability

Mistral AI Observability

Overview

Monitor Mistral AI API usage, latency, token consumption, error rates, and costs. Covers instrumented client wrapper, Prometheus metrics, Grafana dashboard panels, alerting rules, and structured logging.

Prerequisites

• Mistral API integration in production
• Prometheus or OpenTelemetry-compatible metrics backend
• Alerting system (Alertmanager, PagerDuty, or similar)

Instructions

Step 1: Instrumented Client Wrapper

import { Mistral } from '@mistralai/mistralai';

const PRICING: Record<string, { input: number; output: number }> = {
  'mistral-small-latest':  { input: 0.10, output: 0.30 },
  'mistral-large-latest':  { input: 0.50, output: 1.50 },
  'codestral-latest':      { input: 0.30, output: 0.90 },
  'mistral-embed':         { input: 0.10, output: 0 },
};

interface MetricsEvent {
  model: string;
  endpoint: string;
  durationMs: number;
  status: 'success' | 'error';
  statusCode?: number;
  inputTokens?: number;
  outputTokens?: number;
  costUsd?: number;
}

function emitMetrics(event: MetricsEvent): void {
  // Push to your metrics backend (Prometheus, Datadog, etc.)
  console.log(JSON.stringify({ type: 'mistral_metric', ...event }));
}

async function instrumentedChat(
  client: Mistral,
  model: string,
  messages: any[],
  options?: any,
) {
  const start = performance.now();
  try {
    const response = await client.chat.complete({ model, messages, ...options });
    const duration = Math.round(performance.now() - start);
    const pricing = PRICING[model] ?? PRICING['mistral-small-latest'];
    const pt = response.usage?.promptTokens ?? 0;
    const ct = response.usage?.completionTokens ?? 0;

    emitMetrics({
      model,
      endpoint: 'chat.complete',
      durationMs: duration,
      status: 'success',
      inputTokens: pt,
      outputTokens: ct,
      costUsd: (pt / 1e6) * pricing.input + (ct / 1e6) * pricing.output,
    });

    return response;
  } catch (error: any) {
    emitMetrics({
      model,
      endpoint: 'chat.complete',
      durationMs: Math.round(performance.now() - start),
      status: 'error',
      statusCode: error.status,
    });
    throw error;
  }
}

Step 2: Prometheus Metrics

// Using prom-client
import { Counter, Histogram, Gauge } from 'prom-client';

const mistralRequests = new Counter({
  name: 'mistral_requests_total',
  help: 'Total Mistral API requests',
  labelNames: ['model', 'endpoint', 'status'],
});

const mistralDuration = new Histogram({
  name: 'mistral_request_duration_ms',
  help: 'Mistral request duration in milliseconds',
  labelNames: ['model', 'endpoint'],
  buckets: [100, 250, 500, 1000, 2500, 5000, 10000],
});

const mistralTokens = new Counter({
  name: 'mistral_tokens_total',
  help: 'Total tokens consumed',
  labelNames: ['model', 'direction'], // direction: input | output
});

const mistralCost = new Counter({
  name: 'mistral_cost_usd_total',
  help: 'Estimated cost in USD',
  labelNames: ['model'],
});

const mistralErrors = new Counter({
  name: 'mistral_errors_total',
  help: 'Total Mistral errors',
  labelNames: ['model', 'status_code'],
});

// Record metrics from instrumented wrapper
function recordPrometheusMetrics(event: MetricsEvent): void {
  mistralRequests.inc({ model: event.model, endpoint: event.endpoint, status: event.status });
  mistralDuration.observe({ model: event.model, endpoint: event.endpoint }, event.durationMs);

  if (event.status === 'success') {
    if (event.inputTokens) mistralTokens.inc({ model: event.model, direction: 'input' }, event.inputTokens);
    if (event.outputTokens) mistralTokens.inc({ model: event.model, direction: 'output' }, event.outputTokens);
    if (event.costUsd) mistralCost.inc({ model: event.model }, event.costUsd);
  } else {
    mistralErrors.inc({ model: event.model, status_code: String(event.statusCode ?? 'unknown') });
  }
}

Step 3: Alerting Rules

# prometheus/mistral-alerts.yaml
groups:
  - name: mistral
    rules:
      - alert: MistralHighErrorRate
        expr: rate(mistral_errors_total[5m]) / rate(mistral_requests_total[5m]) > 0.05
        for: 5m
        labels: { severity: critical }
        annotations:
          summary: "Mistral error rate exceeds 5%"
          runbook: "See mistral-incident-runbook skill"

      - alert: MistralHighLatency
        expr: histogram_quantile(0.95, rate(mistral_request_duration_ms_bucket[5m])) > 5000
        for: 5m
        labels: { severity: warning }
        annotations:
          summary: "Mistral P95 latency exceeds 5 seconds"

      - alert: MistralRateLimited
        expr: rate(mistral_errors_total{status_code="429"}[5m]) > 0
        for: 2m
        labels: { severity: warning }
        annotations:
          summary: "Mistral rate limiting detected"

      - alert: MistralCostSpike
        expr: increase(mistral_cost_usd_total[1h]) > 10
        labels: { severity: warning }
        annotations:
          summary: "Mistral spend exceeds $10/hour"

      - alert: MistralAuthFailure
        expr: increase(mistral_errors_total{status_code="401"}[5m]) > 0
        labels: { severity: critical }
        annotations:
          summary: "Mistral authentication failing — API key may be revoked"

Step 4: Grafana Dashboard Panels

Key panels to create:

Panel	Query	Type
Request Rate	rate(mistral_requests_total[5m])	Time series
P50/P95/P99 Latency	histogram_quantile(0.95, rate(..._bucket[5m]))	Time series
Token Velocity	rate(mistral_tokens_total{direction="output"}[5m])	Time series
Hourly Cost	increase(mistral_cost_usd_total[1h])	Stat
Error Rate	rate(mistral_errors_total[5m]) by status_code	Time series
Model Distribution	sum by (model) (rate(mistral_requests_total[5m]))	Pie chart

Step 5: Structured Log Format

interface MistralLogEntry {
  ts: string;
  level: 'info' | 'warn' | 'error';
  model: string;
  endpoint: string;
  durationMs: number;
  inputTokens?: number;
  outputTokens?: number;
  costUsd?: number;
  status: string;
  statusCode?: number;
  requestId?: string;
}

function logMistralRequest(entry: MistralLogEntry): void {
  // Ship to SIEM, CloudWatch, or log aggregator
  // NEVER log message content — PII risk
  console.log(JSON.stringify(entry));
}

Error Handling

Issue	Cause	Solution
Missing token counts	Streaming not aggregated	Sum tokens from stream chunks
Cost drift from bill	Pricing table outdated	Update PRICING map when rates change
Alert storm on 429s	Rate limit burst	Tune alert threshold, add request queue
High cardinality	Per-request labels	Never label by request ID or user ID

Resources

• OpenLIT Mistral Monitoring
• Prometheus Client
• Grafana Dashboards

Output

• Instrumented client wrapper with timing and cost tracking
• Prometheus metrics (requests, duration, tokens, cost, errors)
• Alerting rules for error rate, latency, rate limits, cost, auth
• Grafana dashboard panel specifications
• Structured logging format for SIEM integration