Groq Enterprise Access Management

// Create separate keys per team/service via Groq Console Projects
// Each project gets its own API key and can have independent rate limits

// Key naming convention: {team}-{environment}-{purpose}
const KEY_REGISTRY = {
  // Each team gets a separate Groq Project
  "chatbot-prod":         "gsk_...",  // Project: chatbot-production
  "chatbot-staging":      "gsk_...",  // Project: chatbot-staging
  "analytics-prod":       "gsk_...",  // Project: analytics-production
  "batch-processor":      "gsk_...",  // Project: batch-processing
} as const;

// Since Groq keys don't have model scoping, implement it in your gateway
interface TeamConfig {
  allowedModels: string[];
  maxTokensPerRequest: number;
  monthlyBudgetUsd: number;
  rateLimitRPM: number;
}

const TEAM_CONFIGS: Record<string, TeamConfig> = {
  chatbot: {
    allowedModels: ["llama-3.3-70b-versatile", "llama-3.1-8b-instant"],
    maxTokensPerRequest: 2048,
    monthlyBudgetUsd: 200,
    rateLimitRPM: 60,
  },
  analytics: {
    allowedModels: ["llama-3.1-8b-instant"],  // Only cheapest model
    maxTokensPerRequest: 512,
    monthlyBudgetUsd: 50,
    rateLimitRPM: 30,
  },
  research: {
    allowedModels: [
      "llama-3.3-70b-versatile",
      "llama-3.1-8b-instant",
      "meta-llama/llama-4-scout-17b-16e-instruct",
    ],
    maxTokensPerRequest: 4096,
    monthlyBudgetUsd: 500,
    rateLimitRPM: 120,
  },
};

function validateRequest(team: string, model: string, maxTokens: number): void {
  const config = TEAM_CONFIGS[team];
  if (!config) throw new Error(`Unknown team: ${team}`);
  if (!config.allowedModels.includes(model)) {
    throw new Error(`Team ${team} not authorized for model ${model}`);
  }
  if (maxTokens > config.maxTokensPerRequest) {
    throw new Error(`max_tokens ${maxTokens} exceeds limit ${config.maxTokensPerRequest} for team ${team}`);
  }
}

import Groq from "groq-sdk";
import PQueue from "p-queue";

// Per-team rate limiting
const teamQueues = new Map<string, PQueue>();

function getTeamQueue(team: string): PQueue {
  if (!teamQueues.has(team)) {
    const config = TEAM_CONFIGS[team];
    teamQueues.set(team, new PQueue({
      intervalCap: config?.rateLimitRPM || 30,
      interval: 60_000,
      concurrency: 5,
    }));
  }
  return teamQueues.get(team)!;
}

// Gateway function: validates, rate-limits, and proxies to Groq
async function groqGateway(
  team: string,
  messages: any[],
  model: string,
  maxTokens: number
) {
  // Validate permissions
  validateRequest(team, model, maxTokens);

  // Check budget
  const monthlySpend = await getTeamMonthlySpend(team);
  const config = TEAM_CONFIGS[team];
  if (monthlySpend >= config.monthlyBudgetUsd) {
    throw new Error(`Team ${team} monthly budget of $${config.monthlyBudgetUsd} exhausted`);
  }

  // Rate-limited execution
  const queue = getTeamQueue(team);
  return queue.add(async () => {
    const groq = new Groq({ apiKey: getTeamApiKey(team) });
    const result = await groq.chat.completions.create({
      model,
      messages,
      max_tokens: maxTokens,
    });

    // Track usage
    await recordTeamUsage(team, model, result.usage!);
    return result;
  });
}

## Groq Console Setup (per organization)

1. Go to console.groq.com > Organization > Billing
2. Set monthly spending cap
3. Configure alerts at 50%, 80%, 95% thresholds
4. Enable auto-pause when cap is reached

## Application-Level Controls (per team)

// Track spending per team
const teamSpending = new Map<string, number>();

async function recordTeamUsage(
  team: string,
  model: string,
  usage: any
): Promise<void> {
  const pricing: Record<string, { input: number; output: number }> = {
    "llama-3.1-8b-instant": { input: 0.05, output: 0.08 },
    "llama-3.3-70b-versatile": { input: 0.59, output: 0.79 },
    "meta-llama/llama-4-scout-17b-16e-instruct": { input: 0.11, output: 0.34 },
  };

  const price = pricing[model] || { input: 0.10, output: 0.10 };
  const cost =
    (usage.prompt_tokens / 1_000_000) * price.input +
    (usage.completion_tokens / 1_000_000) * price.output;

  const current = teamSpending.get(team) || 0;
  teamSpending.set(team, current + cost);

  // Alert at thresholds
  const budget = TEAM_CONFIGS[team].monthlyBudgetUsd;
  const pct = ((current + cost) / budget) * 100;

  if (pct >= 95) {
    console.error(`[ALERT] Team ${team} at ${pct.toFixed(0)}% of monthly budget!`);
  } else if (pct >= 80) {
    console.warn(`[WARN] Team ${team} at ${pct.toFixed(0)}% of monthly budget`);
  }
}

set -euo pipefail
# Zero-downtime key rotation process:

# 1. Create new key in Groq Console (same Project)
#    Name: chatbot-prod-2026-04

# 2. Deploy new key alongside old key
#    Both keys are valid simultaneously

# 3. Update secret manager
#    AWS: aws secretsmanager update-secret --secret-id groq/chatbot-prod --secret-string "gsk_new_..."
#    GCP: echo -n "gsk_new_..." | gcloud secrets versions add groq-chatbot-prod --data-file=-

# 4. Restart services to pick up new key

# 5. Monitor for 24h -- verify no requests on old key

# 6. Delete old key in Groq Console

// Weekly usage report per team
function weeklyReport(records: Array<{ team: string; model: string; cost: number; tokens: number }>) {
  const byTeam: Record<string, { cost: number; tokens: number; topModel: string }> = {};

  for (const r of records) {
    if (!byTeam[r.team]) byTeam[r.team] = { cost: 0, tokens: 0, topModel: "" };
    byTeam[r.team].cost += r.cost;
    byTeam[r.team].tokens += r.tokens;
  }

  console.table(
    Object.entries(byTeam).map(([team, data]) => ({
      team,
      cost: `$${data.cost.toFixed(2)}`,
      tokens: data.tokens.toLocaleString(),
      budget: `$${TEAM_CONFIGS[team]?.monthlyBudgetUsd || "N/A"}`,
    }))
  );
}