ideogram-observability（表意文字可观察性）

Ideogram Observability

Overview

Monitor Ideogram AI image generation for latency, cost, error rates, and content safety rejections. Key metrics: generation duration (5-25s depending on model), credit burn rate, safety filter rejection rate, and API availability. Ideogram's API is synchronous, so all observability is request-level instrumentation.

Key Metrics

Instructions

Step 1: Instrumented Generation Wrapper

import { performance } from "perf_hooks";

interface GenerationMetrics {
  duration: number;
  model: string;
  style: string;
  status: "success" | "error" | "safety_rejected" | "rate_limited";
  seed?: number;
  resolution?: string;
}

const metricsLog: GenerationMetrics[] = [];

async function instrumentedGenerate(
  prompt: string,
  options: { model?: string; style_type?: string; aspect_ratio?: string } = {}
) {
  const model = options.model ?? "V_2";
  const style = options.style_type ?? "AUTO";
  const start = performance.now();

  try {
    const response = await fetch("https://api.ideogram.ai/generate", {
      method: "POST",
      headers: {
        "Api-Key": process.env.IDEOGRAM_API_KEY!,
        "Content-Type": "application/json",
      },
      body: JSON.stringify({
        image_request: { prompt, model, style_type: style, ...options, magic_prompt_option: "AUTO" },
      }),
    });

    const duration = performance.now() - start;

    if (response.status === 422) {
      recordMetric({ duration, model, style, status: "safety_rejected" });
      throw new Error("Safety filter rejected prompt");
    }
    if (response.status === 429) {
      recordMetric({ duration, model, style, status: "rate_limited" });
      throw new Error("Rate limited");
    }
    if (!response.ok) {
      recordMetric({ duration, model, style, status: "error" });
      throw new Error(`API error: ${response.status}`);
    }

    const result = await response.json();
    const image = result.data[0];

    recordMetric({
      duration, model, style, status: "success",
      seed: image.seed, resolution: image.resolution,
    });

    return result;
  } catch (err) {
    if (!metricsLog.find(m => m.duration === performance.now() - start)) {
      recordMetric({ duration: performance.now() - start, model, style, status: "error" });
    }
    throw err;
  }
}

function recordMetric(metric: GenerationMetrics) {
  metricsLog.push(metric);

  // Emit to your metrics backend
  console.log(JSON.stringify({
    event: "ideogram.generation",
    ...metric,
    timestamp: new Date().toISOString(),
  }));
}

Step 2: Cost Estimation Metrics

const MODEL_COST_USD: Record<string, number> = {
  V_2_TURBO: 0.05, V_2: 0.08, V_2A: 0.04, V_2A_TURBO: 0.025,
};

function estimateCost(model: string, numImages: number = 1): number {
  return (MODEL_COST_USD[model] ?? 0.08) * numImages;
}

function costReport(metrics: GenerationMetrics[]) {
  const successful = metrics.filter(m => m.status === "success");
  const totalCost = successful.reduce((sum, m) => sum + estimateCost(m.model), 0);
  const byModel = Object.groupBy(successful, m => m.model);

  console.log("=== Ideogram Cost Report ===");
  console.log(`Total generations: ${successful.length}`);
  console.log(`Estimated cost: $${totalCost.toFixed(2)}`);

  for (const [model, gens] of Object.entries(byModel)) {
    const cost = (gens?.length ?? 0) * (MODEL_COST_USD[model] ?? 0.08);
    console.log(`  ${model}: ${gens?.length ?? 0} images, ~$${cost.toFixed(2)}`);
  }
}

Step 3: Prometheus Metrics (Optional)

import { Counter, Histogram, register } from "prom-client";

const generationDuration = new Histogram({
  name: "ideogram_generation_duration_seconds",
  help: "Ideogram image generation duration",
  labelNames: ["model", "style", "status"],
  buckets: [2, 5, 10, 15, 20, 30, 60],
});

const generationTotal = new Counter({
  name: "ideogram_generations_total",
  help: "Total Ideogram generations",
  labelNames: ["model", "status"],
});

const estimatedCostTotal = new Counter({
  name: "ideogram_estimated_cost_usd",
  help: "Estimated Ideogram API cost in USD",
  labelNames: ["model"],
});

// Expose metrics endpoint
app.get("/metrics", async (req, res) => {
  res.set("Content-Type", register.contentType);
  res.end(await register.metrics());
});

Step 4: Alerting Rules

# prometheus-rules.yml
groups:
  - name: ideogram
    rules:
      - alert: IdeogramGenerationSlow
        expr: histogram_quantile(0.95, rate(ideogram_generation_duration_seconds_bucket[15m])) > 25
        for: 5m
        annotations:
          summary: "Ideogram P95 generation time exceeds 25 seconds"

      - alert: IdeogramHighErrorRate
        expr: rate(ideogram_generations_total{status="error"}[10m]) / rate(ideogram_generations_total[10m]) > 0.05
        for: 5m
        annotations:
          summary: "Ideogram error rate exceeds 5%"

      - alert: IdeogramHighCostRate
        expr: rate(ideogram_estimated_cost_usd[1h]) > 10
        annotations:
          summary: "Ideogram burning >$10/hour"

      - alert: IdeogramSafetyRejectionSpike
        expr: rate(ideogram_generations_total{status="safety_rejected"}[1h]) / rate(ideogram_generations_total[1h]) > 0.1
        annotations:
          summary: "Ideogram safety rejection rate exceeds 10%"

Step 5: Dashboard Panel Queries

# Grafana dashboard panels:
# 1. Generation volume:     sum(rate(ideogram_generations_total[5m])) by (model)
# 2. Latency distribution:  histogram_quantile(0.5, rate(ideogram_generation_duration_seconds_bucket[5m]))
# 3. Error rate:            sum(rate(ideogram_generations_total{status!="success"}[5m])) / sum(rate(ideogram_generations_total[5m]))
# 4. Cost per hour:         sum(rate(ideogram_estimated_cost_usd[1h]))
# 5. Safety rejections:     sum(rate(ideogram_generations_total{status="safety_rejected"}[1h]))

Error Handling

Output

• Instrumented generation wrapper with metrics collection
• Cost estimation and reporting
• Prometheus metrics with alerting rules
• Grafana dashboard query templates

Resources

• Ideogram API Overview
• Prometheus Client
• Grafana Dashboards

Next Steps

For incident response, see ideogram-incident-runbook.