Cohere Rate Limits

Cohere Rate Limits

Overview

Handle Cohere rate limits with exponential backoff, request queuing, and proactive throttling. Real rate limits from Cohere's documentation.

Prerequisites

• cohere-ai SDK installed
• Understanding of async/await patterns

Actual Cohere Rate Limits

Trial keys are free. Production keys require billing at dashboard.cohere.com.

Instructions

Step 1: Exponential Backoff with Jitter

import { CohereError, CohereTimeoutError } from 'cohere-ai';

interface RetryConfig {
  maxRetries: number;
  baseDelayMs: number;
  maxDelayMs: number;
}

const DEFAULT_RETRY: RetryConfig = {
  maxRetries: 5,
  baseDelayMs: 1000,
  maxDelayMs: 60_000,
};

async function withBackoff<T>(
  operation: () => Promise<T>,
  config = DEFAULT_RETRY
): Promise<T> {
  for (let attempt = 0; attempt <= config.maxRetries; attempt++) {
    try {
      return await operation();
    } catch (err) {
      if (attempt === config.maxRetries) throw err;

      // Only retry on rate limits (429) and server errors (5xx)
      let shouldRetry = false;
      let retryAfterMs: number | undefined;

      if (err instanceof CohereError) {
        if (err.statusCode === 429) {
          shouldRetry = true;
          // Cohere returns Retry-After header (seconds)
          // SDK may expose this via err.headers
        } else if (err.statusCode && err.statusCode >= 500) {
          shouldRetry = true;
        }
      } else if (err instanceof CohereTimeoutError) {
        shouldRetry = true;
      }

      if (!shouldRetry) throw err;

      // Exponential delay with jitter
      const exponential = config.baseDelayMs * Math.pow(2, attempt);
      const jitter = Math.random() * config.baseDelayMs;
      const delay = Math.min(exponential + jitter, config.maxDelayMs);

      console.warn(`Cohere retry ${attempt + 1}/${config.maxRetries} in ${delay.toFixed(0)}ms`);
      await new Promise(r => setTimeout(r, retryAfterMs ?? delay));
    }
  }
  throw new Error('Unreachable');
}

Step 2: Request Queue (Concurrency-Limited)

import PQueue from 'p-queue';

// Match rate limits: trial=20/min for chat, production=1000/min
function createCohereQueue(callsPerMinute: number) {
  return new PQueue({
    concurrency: 5,
    interval: 60_000,
    intervalCap: callsPerMinute,
  });
}

// Trial key queue
const trialChatQueue = createCohereQueue(20);
const trialEmbedQueue = createCohereQueue(5);

// Production key queue
const prodQueue = createCohereQueue(1000);

// Usage
async function queuedChat(params: any) {
  return trialChatQueue.add(() =>
    withBackoff(() => cohere.chat(params))
  );
}

Step 3: Proactive Rate Tracking

class RateLimitTracker {
  private windows: Map<string, number[]> = new Map();

  constructor(private limitsPerMinute: Record<string, number>) {}

  canProceed(endpoint: string): boolean {
    const limit = this.limitsPerMinute[endpoint] ?? 1000;
    const now = Date.now();
    const window = this.windows.get(endpoint) ?? [];

    // Remove entries older than 1 minute
    const active = window.filter(t => now - t < 60_000);
    this.windows.set(endpoint, active);

    return active.length < limit;
  }

  record(endpoint: string): void {
    const window = this.windows.get(endpoint) ?? [];
    window.push(Date.now());
    this.windows.set(endpoint, window);
  }

  waitTime(endpoint: string): number {
    const limit = this.limitsPerMinute[endpoint] ?? 1000;
    const window = this.windows.get(endpoint) ?? [];
    const now = Date.now();
    const active = window.filter(t => now - t < 60_000);

    if (active.length < limit) return 0;
    return 60_000 - (now - active[0]); // Wait until oldest entry expires
  }
}

// Trial key tracker
const tracker = new RateLimitTracker({
  chat: 20,
  embed: 5,
  rerank: 5,
  classify: 5,
});

// Use before each call
async function trackedEmbed(params: any) {
  const wait = tracker.waitTime('embed');
  if (wait > 0) {
    console.log(`Throttling embed: waiting ${wait}ms`);
    await new Promise(r => setTimeout(r, wait));
  }
  tracker.record('embed');
  return withBackoff(() => cohere.embed(params));
}

Step 4: Batch-Aware Embedding

// Embed supports up to 96 texts per call — maximize batch size to reduce calls
async function efficientEmbed(
  texts: string[],
  inputType: 'search_document' | 'search_query' = 'search_document'
): Promise<number[][]> {
  const BATCH_SIZE = 96; // Cohere max per request
  const allVectors: number[][] = [];

  for (let i = 0; i < texts.length; i += BATCH_SIZE) {
    const batch = texts.slice(i, i + BATCH_SIZE);

    const response = await trackedEmbed({
      model: 'embed-v4.0',
      texts: batch,
      inputType,
      embeddingTypes: ['float'],
    });

    allVectors.push(...response.embeddings.float);
  }

  return allVectors;
}

// 960 texts = 10 API calls (not 960)
const vectors = await efficientEmbed(largeTextArray);

Cost-Aware Rate Limiting

For production keys, rate limits are per-minute but costs are per-token:

class TokenBudget {
  private tokensUsed = 0;
  private readonly resetInterval: NodeJS.Timer;

  constructor(
    private maxTokensPerMinute: number,
    private alertCallback?: (used: number) => void
  ) {
    // Reset every minute
    this.resetInterval = setInterval(() => { this.tokensUsed = 0; }, 60_000);
  }

  canAfford(estimatedTokens: number): boolean {
    return this.tokensUsed + estimatedTokens <= this.maxTokensPerMinute;
  }

  record(actualTokens: number): void {
    this.tokensUsed += actualTokens;
    if (this.tokensUsed > this.maxTokensPerMinute * 0.8) {
      this.alertCallback?.(this.tokensUsed);
    }
  }

  dispose(): void {
    clearInterval(this.resetInterval);
  }
}

Output

• Automatic retry with exponential backoff + jitter
• Concurrency-limited request queue matching Cohere rate limits
• Proactive throttling before hitting limits
• Batch-optimized embedding to minimize API calls

Error Handling

Resources

• Cohere Rate Limits
• Cohere API Keys
• p-queue

Next Steps

For security configuration, see cohere-security-basics.

Source: jeremylongshore/claude-code-plugins-plus-skills → plugins/saas-packs/cohere-pack/skills/cohere-rate-limits/SKILL.md