Cohere Tool Use & Agents (Core Workflow B)

Overview

Build multi-step tool-using agents with Cohere's Chat API v2. The model decides which tools to call, you execute them, and feed results back in a loop until the task is complete.

Prerequisites

• Completed cohere-install-auth setup
• Understanding of cohere-core-workflow-a (RAG)
• Command R7B or newer model (required for tool use)

Instructions

Step 1: Define Tools

import { CohereClientV2 } from 'cohere-ai';

const cohere = new CohereClientV2();

// Define tools the model can call
const tools = [
  {
    type: 'function' as const,
    function: {
      name: 'get_weather',
      description: 'Get current weather for a city',
      parameters: {
        type: 'object' as const,
        properties: {
          city: { type: 'string', description: 'City name' },
          unit: { type: 'string', enum: ['celsius', 'fahrenheit'], description: 'Temperature unit' },
        },
        required: ['city'],
      },
    },
  },
  {
    type: 'function' as const,
    function: {
      name: 'search_database',
      description: 'Search internal database for records',
      parameters: {
        type: 'object' as const,
        properties: {
          query: { type: 'string', description: 'Search query' },
          limit: { type: 'number', description: 'Max results' },
        },
        required: ['query'],
      },
    },
  },
];

Step 2: Implement Tool Executors

// Map tool names to actual implementations
const toolExecutors: Record<string, (args: any) => Promise<string>> = {
  get_weather: async ({ city, unit = 'celsius' }) => {
    // Replace with real weather API call
    return JSON.stringify({
      city,
      temperature: unit === 'celsius' ? 22 : 72,
      unit,
      condition: 'partly cloudy',
    });
  },

  search_database: async ({ query, limit = 5 }) => {
    // Replace with real database query
    return JSON.stringify({
      results: [
        { id: 1, title: `Result for: ${query}`, relevance: 0.95 },
      ],
      total: 1,
    });
  },
};

Step 3: Single-Step Tool Use

async function singleStepToolUse(userMessage: string) {
  // 1. Send message with tools
  const response = await cohere.chat({
    model: 'command-a-03-2025',
    messages: [{ role: 'user', content: userMessage }],
    tools,
  });

  // 2. Check if model wants to call tools
  if (response.finishReason === 'TOOL_CALL') {
    const toolCalls = response.message?.toolCalls ?? [];

    // 3. Execute each tool call
    const toolResults = a

Cohere Cost Tuning

Overview

Optimize Cohere costs through model selection, token budgets, embedding compression, and usage monitoring. Cohere pricing is token-based with separate input/output rates.

Prerequisites

• Cohere production key (trial is free but limited)
• Access to dashboard.cohere.com billing page

Cohere Pricing Model

Key principle: Cohere charges per token. Input tokens and output tokens have different rates. Embed, Rerank, and Classify have separate pricing based on search units.

Model Cost Comparison

Non-Chat Pricing

Instructions

Strategy 1: Model Tiering

type CostTier = 'economy' | 'standard' | 'premium';

function selectModel(tier: CostTier): string {
  switch (tier) {
    case 'economy':  return 'command-r7b-12-2024';    // ~5x cheaper
    case 'standard': return 'command-r-08-2024';       // Good balance
    case 'premium':  return 'command-a-03-2025';       // Best quality
  }
}

// Route by use case
function routeModel(task: string): string {
  // High-volume, simple tasks → cheapest model
  if (['classify', 'extract', 'summarize-short'].includes(task)) {
    return selectModel('economy');
  }
  // RAG, moderate complexity
  if (['rag', 'search', 'qa'].includes(task)) {
    return selectModel('standard');
  }
  // Complex reasoning, user-facing
  return selectModel('premium');
}

Strate

Cohere Debug Bundle

Overview

Collect all diagnostic information needed to resolve Cohere API v2 issues. Generates a support-ready bundle with environment info, request/response logs, and SDK version data.

Prerequisites

• cohere-ai SDK installed
• Access to application logs
• curl and jq available

Instructions

Step 1: Create Debug Bundle Script

#!/bin/bash
# cohere-debug-bundle.sh
set -euo pipefail

BUNDLE_DIR="cohere-debug-$(date +%Y%m%d-%H%M%S)"
mkdir -p "$BUNDLE_DIR"

echo "=== Cohere Debug Bundle ===" > "$BUNDLE_DIR/summary.txt"
echo "Generated: $(date -u +%Y-%m-%dT%H:%M:%SZ)" >> "$BUNDLE_DIR/summary.txt"
echo "" >> "$BUNDLE_DIR/summary.txt"

Step 2: Collect Environment and SDK Info

# Runtime versions
echo "--- Runtime ---" >> "$BUNDLE_DIR/summary.txt"
node --version >> "$BUNDLE_DIR/summary.txt" 2>&1 || echo "Node.js: not found" >> "$BUNDLE_DIR/summary.txt"
python3 --version >> "$BUNDLE_DIR/summary.txt" 2>&1 || echo "Python: not found" >> "$BUNDLE_DIR/summary.txt"

# SDK version
echo "--- SDK ---" >> "$BUNDLE_DIR/summary.txt"
npm list cohere-ai 2>/dev/null >> "$BUNDLE_DIR/summary.txt" || echo "cohere-ai: not installed (npm)" >> "$BUNDLE_DIR/summary.txt"
pip show cohere 2>/dev/null | grep Version >> "$BUNDLE_DIR/summary.txt" || echo "cohere: not installed (pip)" >> "$BUNDLE_DIR/summary.txt"

# API key status (never log the actual key)
echo "--- Auth ---" >> "$BUNDLE_DIR/summary.txt"
if [ -n "${CO_API_KEY:-}" ]; then
  echo "CO_API_KEY: SET (${#CO_API_KEY} chars, starts with ${CO_API_KEY:0:4}...)" >> "$BUNDLE_DIR/summary.txt"
else
  echo "CO_API_KEY: NOT SET" >> "$BUNDLE_DIR/summary.txt"
fi

Step 3: Test API Connectivity

echo "--- API Connectivity ---" >> "$BUNDLE_DIR/summary.txt"

# Test each endpoint
for endpoint in chat embed rerank classify; do
  STATUS=$(curl -s -o /dev/null -w "%{http_code}" \
    -X POST "https://api.cohere.com/v2/$endpoint" \
    -H "Authorization: Bearer ${CO_API_KEY:-invalid}" \
    -H "Content-Type: application/json" \
    -d '{}' 2>/dev/null || echo "UNREACHABLE")
  echo "$endpoint: HTTP $STATUS" >> "$BUNDLE_DIR/summary.txt"
done

# Check service status
echo "" >> "$BUNDLE_DIR/summary.txt"
echo "--- Service Stat

Cohere Deploy Integration

Overview

Deploy Cohere API v2 applications to Vercel, Fly.io, and Google Cloud Run with proper secrets management and health checks.

Prerequisites

• Cohere production API key (not trial)
• Platform CLI installed (vercel, fly, or gcloud)
• Application tested locally with real API calls

Instructions

Vercel Deployment

# Add Cohere API key as Vercel environment variable
vercel env add CO_API_KEY production
# Paste your production key when prompted

# Deploy
vercel --prod

vercel.json:

{
  "env": {
    "CO_API_KEY": "@co_api_key"
  },
  "functions": {
    "api/**/*.ts": {
      "maxDuration": 30
    }
  }
}

Vercel API Route (streaming chat):

// api/chat/route.ts
import { CohereClientV2 } from 'cohere-ai';

const cohere = new CohereClientV2();

export async function POST(req: Request) {
  const { message } = await req.json();

  const stream = await cohere.chatStream({
    model: 'command-a-03-2025',
    messages: [{ role: 'user', content: message }],
  });

  const encoder = new TextEncoder();
  const readable = new ReadableStream({
    async start(controller) {
      for await (const event of stream) {
        if (event.type === 'content-delta') {
          const text = event.delta?.message?.content?.text ?? '';
          controller.enqueue(encoder.encode(`data: ${JSON.stringify({ text })}\n\n`));
        }
      }
      controller.enqueue(encoder.encode('data: [DONE]\n\n'));
      controller.close();
    },
  });

  return new Response(readable, {
    headers: { 'Content-Type': 'text/event-stream', 'Cache-Control': 'no-cache' },
  });
}

Fly.io Deployment

# Set Cohere API key
fly secrets set CO_API_KEY="your-production-key"

# Deploy
fly deploy

fly.toml:

app = "my-cohere-app"
primary_region = "iad"

[env]
  NODE_ENV = "production"

[http_service]
  internal_port = 3000
  force_https = true
  auto_stop_machines = true
  auto_start_machines = true
  min_machines_running = 1

[[services.http_checks]]
  interval = 30000
  timeout = 5000
  path = "/api/health"

Google Cloud Run Deployment

#!/bin/bash
PROJECT_ID="${GOOGLE_CLOUD_PROJECT}"
SERVICE="cohere-app"
REGION="us-central1"

# Store key in Secret Manager
echo -n "$CO_API_KEY" | gcloud secrets create cohere-api-key --data-file=-

# Build and deploy
gcloud builds submit --tag gcr.io/$PROJECT_ID/$SERV

Cohere Enterprise RBAC

Overview

Configure enterprise-grade access control for Cohere API v2 with multi-team API key management, per-team model/budget restrictions, and audit trails.

Prerequisites

• Cohere production API keys
• Understanding of your team/service structure
• Secret management infrastructure

Cohere Access Model

Cohere uses API key-based access control (no built-in RBAC or SSO). Enterprise patterns are implemented in your application layer.

Instructions

Step 1: Multi-Team Key Strategy

// Each team gets their own API key for tracking and revocation
interface TeamConfig {
  name: string;
  apiKeyEnvVar: string;
  allowedModels: string[];
  maxTokensPerCall: number;
  dailyBudgetUSD: number;
}

const teamConfigs: Record<string, TeamConfig> = {
  search: {
    name: 'Search Team',
    apiKeyEnvVar: 'CO_API_KEY_SEARCH',
    allowedModels: ['embed-v4.0', 'rerank-v3.5', 'command-r-08-2024'],
    maxTokensPerCall: 1000,
    dailyBudgetUSD: 50,
  },
  chatbot: {
    name: 'Chatbot Team',
    apiKeyEnvVar: 'CO_API_KEY_CHATBOT',
    allowedModels: ['command-a-03-2025', 'command-r7b-12-2024'],
    maxTokensPerCall: 4096,
    dailyBudgetUSD: 200,
  },
  ml: {
    name: 'ML Team',
    apiKeyEnvVar: 'CO_API_KEY_ML',
    allowedModels: ['embed-v4.0', 'embed-multilingual-v3.0'],
    maxTokensPerCall: 500,
    dailyBudgetUSD: 100,
  },
};

Step 2: Team-Scoped Client Factory

import { CohereClientV2 } from 'cohere-ai';

const clients = new Map<string, CohereClientV2>();

export function getCohereForTeam(teamId: string): CohereClientV2 {
  if (!clients.has(teamId)) {
    const config = teamConfigs[teamId];
    if (!config) throw new Error(`Unknown team: ${teamId}`);

    const apiKey = process.env[config.apiKeyEnvVar];
    if (!apiKey) throw new Error(`${config.apiKeyEnvVar} not set for team ${teamId}`);

    clients.set(teamId, new CohereClientV2({ token: apiKey }));
  }
  return clients.get(teamId)!;
}

Step 3: Model Access Enforcement

function enforceModelAccess(teamId: string, requestedModel: string): void {
  const config = teamConfigs[teamId];
  if (!config) throw new Error(`Unknown team: ${teamId

Cohere Hello World

Overview

Three minimal working examples: Chat completion, text embedding, and search reranking. Each demonstrates a core Cohere API v2 endpoint.

Prerequisites

• Completed cohere-install-auth setup
• cohere-ai package installed
• COAPIKEY environment variable set

Instructions

Example 1: Chat Completion

import { CohereClientV2 } from 'cohere-ai';

const cohere = new CohereClientV2();

async function chat() {
  const response = await cohere.chat({
    model: 'command-a-03-2025',
    messages: [
      { role: 'system', content: 'You are a helpful coding assistant.' },
      { role: 'user', content: 'Explain what a closure is in JavaScript in 2 sentences.' },
    ],
  });

  console.log(response.message?.content?.[0]?.text);
}

chat().catch(console.error);

Example 2: Text Embedding

async function embed() {
  const response = await cohere.embed({
    model: 'embed-v4.0',
    texts: ['Cohere builds enterprise AI', 'LLMs power modern search'],
    inputType: 'search_document',
    embeddingTypes: ['float'],
  });

  const vectors = response.embeddings.float;
  console.log(`Generated ${vectors.length} embeddings`);
  console.log(`Dimensions: ${vectors[0].length}`);
}

embed().catch(console.error);

Example 3: Search Reranking

async function rerank() {
  const response = await cohere.rerank({
    model: 'rerank-v3.5',
    query: 'What is machine learning?',
    documents: [
      'Machine learning is a subset of artificial intelligence.',
      'The weather today is sunny and warm.',
      'Deep learning uses neural networks with many layers.',
      'I enjoy cooking Italian food on weekends.',
    ],
    topN: 2,
  });

  for (const result of response.results) {
    console.log(`[${result.relevanceScore.toFixed(3)}] ${result.index}`);
  }
}

rerank().catch(console.error);

Example 4: Streaming Chat

async function streamChat() {
  const stream = await cohere.chatStream({
    model: 'command-a-03-2025',
    messages: [
      { role: 'user', content: 'Write a haiku about APIs.' },
    ],
  });

  for await (const event of stream) {
    if (event.type === 'content-delta') {
      process.stdout.write(event.delta?.message?.content?.text ?? '');
    }
  }
  console.log(); // newline
}

streamChat().catch(console.error);

Python Equivalents

import cohere

co = cohere.ClientV2()

# Chat
response = co.chat(
    model="command-a-03-2025",
    messages=[{"role": "us

Cohere Incident Runbook

Overview

Rapid incident response procedures for Cohere API v2 outages. Covers triage, mitigation, communication, and postmortem for Chat, Embed, Rerank, and Classify endpoints.

Prerequisites

• Access to status.cohere.com
• kubectl access to production cluster
• Prometheus/Grafana access
• PagerDuty/Slack communication channels

Severity Levels

Quick Triage (Run These First)

# 1. Check Cohere service status
curl -s https://status.cohere.com/api/v2/status.json | jq '.status.description'

# 2. Test each endpoint directly
echo "--- Chat ---"
curl -s -o /dev/null -w "%{http_code}" \
  -X POST https://api.cohere.com/v2/chat \
  -H "Authorization: Bearer $CO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"command-r7b-12-2024","messages":[{"role":"user","content":"ping"}]}'

echo -e "\n--- Embed ---"
curl -s -o /dev/null -w "%{http_code}" \
  -X POST https://api.cohere.com/v2/embed \
  -H "Authorization: Bearer $CO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"embed-v4.0","texts":["test"],"input_type":"search_document","embedding_types":["float"]}'

echo -e "\n--- Rerank ---"
curl -s -o /dev/null -w "%{http_code}" \
  -X POST https://api.cohere.com/v2/rerank \
  -H "Authorization: Bearer $CO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"rerank-v3.5","query":"test","documents":["a","b"]}'

# 3. Check our app health
curl -sf https://api.yourapp.com/api/health | jq '.cohere'

# 4. Check error rate (last 5 min)
curl -s "localhost:9090/api/v1/query?query=rate(cohere_errors_total[5m])" | jq '.data.result'

Decision Tree

Cohere API returning errors?
├─ YES: Is status.cohere.com showing incident?
│   ├─ YES → Cohere-side outage. Enable fallback. Monitor status page.
│   └─ NO → Check our API key and configuration.
│       ├─

Cohere Install & Auth

Overview

Set up the Cohere SDK (v2) and configure authentication for Chat, Embed, Rerank, and Classify endpoints.

Prerequisites

• Node.js 18+ or Python 3.10+
• Package manager (npm, pnpm, or pip)
• Cohere account at dashboard.cohere.com
• API key from Cohere dashboard (trial keys are free, production keys require billing)

Instructions

Step 1: Install SDK

# Node.js / TypeScript
npm install cohere-ai

# Python
pip install cohere

Step 2: Configure API Key

# Set environment variable
export CO_API_KEY="your-api-key-here"

# Or create .env file (add .env to .gitignore!)
echo 'CO_API_KEY=your-api-key-here' >> .env

Key types:

• Trial key — free, rate-limited (5-20 calls/min per endpoint, 1000/month others)
• Production key — metered billing, 1000 calls/min all endpoints, unlimited monthly

Step 3: Verify Connection (TypeScript)

import { CohereClientV2 } from 'cohere-ai';

const cohere = new CohereClientV2({
  token: process.env.CO_API_KEY,
});

async function verify() {
  const response = await cohere.chat({
    model: 'command-a-03-2025',
    messages: [
      { role: 'user', content: 'Say "connection verified" and nothing else.' },
    ],
  });
  console.log('Status:', response.message?.content?.[0]?.text);
}

verify().catch(console.error);

Step 4: Verify Connection (Python)

import cohere
import os

co = cohere.ClientV2(api_key=os.environ.get("CO_API_KEY"))

response = co.chat(
    model="command-a-03-2025",
    messages=[
        {"role": "user", "content": "Say 'connection verified' and nothing else."}
    ],
)
print("Status:", response.message.content[0].text)

Available Models

my-cohere-project/
├── src/
│   ├── cohere/
│   │   ├── client.ts       # CohereClientV2 wrapper
│   │   ├── chat.ts         # Chat completions
│   │   ├── embed.ts        # Embedding operations
│   │   └── rerank.ts       # Reranking operations
│   └── index.ts
├── tests/
│   ├── chat.test.ts
│   ├── embed.test.ts
│   └── fixtures/
│       └── responses.ts    # Mock API responses
├── .env.local              # Local secrets (git-ignored)
├── .env.example            # Template for team
└── package.json

{
  "scripts": {
    "dev": "tsx watch src/index.ts",
    "test": "vitest",
    "test:watch": "vitest --watch",
    "test:integration": "COHERE_INTEGRATION=1 vitest --run"
  },
  "dependencies": {
    "cohere-ai": "^7.0.0"
  },
  "devDependencies": {
    "tsx": "^4.0.0",
    "vitest": "^2.0.0",
    "typescript": "^5.5.0"
  }
}

// src/cohere/client.ts
import { CohereClientV2 } from 'cohere-ai';

let instance: CohereClientV2 | null = null;

export function getCohere(): CohereClientV2 {
  if (!instance) {
    instance = new CohereClientV2({
      token: process.env.CO_API_KEY,
    });
  }
  return instance;
}

// Reset for testing
export function resetClient(): void {
  instance = null;
}

// tests/fixtures/responses.ts
export const mockChatResponse = {
  id: 'test-chat-id',
  message: {
    role: 'assistant' as const,
    content: [{ type: 'text' as const, text: 'Mocked response' }],
  },
  finishReason: 'COMPLETE' as const,
  usage: { billedUnits: { inputTokens: 10, outputTokens: 5 } },
};

export const mockEmbedResponse = {
  id: 'test-embed-id',
  embeddings: {
    float: [[0.1, 0.2, 0.3, 0.4]], // truncated for dev
  },
  meta: { billedUnits: { inputTokens: 4 } },
};

export const mockRerankResponse = {
  id: 'test-rerank-id',
  results: [
    { index: 0, relevanceScore: 0.95 },
    { index: 2, relevanceScore: 0.72 },
  ],
  meta: { billedUnits: { searchUnits: 1 } },
};

// tests/chat.test.ts
import { describe, it, expect, vi, beforeEach } from 'vitest&

// --- OpenAI (before) ---
import OpenAI from 'openai';
const openai = new OpenAI();

const response = await openai.chat.completions.create({
  model: 'gpt-4o',
  messages: [
    { role: 'system', content: 'You are helpful.' },
    { role: 'user', content: 'Hello' },
  ],
  max_tokens: 500,
  temperature: 0.7,
});
const text = response.choices[0].message.content;

// --- Cohere (after) ---
import { CohereClientV2 } from 'cohere-ai';
const cohere = new CohereClientV2();

const response = await cohere.chat({
  model: 'command-a-03-2025',   // GPT-4o equivalent
  messages: [
    { role: 'system', content: 'You are helpful.' },  // Same format!
    { role: 'user', content: 'Hello' },
  ],
  maxTokens: 500,               // camelCase, not snake_case
  temperature: 0.7,
});
const text = response.message?.content?.[0]?.text;  // Different response shape

// OpenAI embeddings: 3072 dims (text-embedding-3-large)
// Cohere embeddings: 1024 dims (embed-v4.0)
// IMPORTANT: You CANNOT mix embeddings from different models in the same vector DB

// Migration plan:
// 1. Create new vector collection with Cohere dimensions
// 2. Re-embed all documents with Cohere
// 3. Switch queries to new collection
// 4. Delete old collection

async function migrateEmbeddings(
  documents: Array<{ id: string; text: string }>,
  batchSize = 96
) {
  const cohere = new CohereClientV2();
  let processed = 0;

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

    const response = await cohere.embed({
      model: 'embed-v4.0',
      texts: batch.map(d => d.text),
      inputType: 'search_document',
      embedding

// src/config/cohere.ts
type Environment = 'development' | 'staging' | 'production';

interface CohereEnvConfig {
  chatModel: string;
  embedModel: string;
  rerankModel: string;
  maxTokens: number;
  cacheEnabled: boolean;
  cacheTtlMs: number;
  retries: number;
  timeoutSeconds: number;
}

const configs: Record<Environment, CohereEnvConfig> = {
  development: {
    chatModel: 'command-r7b-12-2024',    // Fastest, cheapest
    embedModel: 'embed-v4.0',
    rerankModel: 'rerank-v3.5',
    maxTokens: 200,                       // Limit for dev
    cacheEnabled: false,                   // See real responses
    cacheTtlMs: 0,
    retries: 1,                            // Fail fast in dev
    timeoutSeconds: 30,
  },
  staging: {
    chatModel: 'command-r-08-2024',       // Mid-tier for testing
    embedModel: 'embed-v4.0',
    rerankModel: 'rerank-v3.5',
    maxTokens: 1000,
    cacheEnabled: true,
    cacheTtlMs: 5 * 60 * 1000,           // 5 minutes
    retries: 3,
    timeoutSeconds: 60,
  },
  production: {
    chatModel: 'command-a-03-2025',       // Best quality
    embedModel: 'embed-v4.0',
    rerankModel: 'rerank-v3.5',
    maxTokens: 4096,
    cacheEnabled: true,
    cacheTtlMs: 15 * 60 * 1000,          // 15 minutes
    retries: 5,
    timeoutSeconds: 120,
  },
};

function detectEnvironment(): Environment {
  const env = process.env.NODE_ENV ?? 'development';
  if (['development', 'staging', 'production'].includes(env)) {
    return env as Environment;
  }
  return 'development';
}

export function getCohereConfig(): CohereEnvConfig & { environment: Environment } {
  const env = detectEnvironment();
  return { ...configs[env], environment: env };
}

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

const registry = new Registry();

// Per-endpoint request counter
const requestCounter = new Counter({
  name: 'cohere_requests_total',
  help: 'Total Cohere API requests',
  labelNames: ['endpoint', 'model', 'status'],
  registers: [registry],
});

// Latency histogram
const requestDuration = new Histogram({
  name: 'cohere_request_duration_seconds',
  help: 'Cohere request duration',
  labelNames: ['endpoint', 'model'],
  buckets: [0.1, 0.25, 0.5, 1, 2.5, 5, 10, 30],
  registers: [registry],
});

// Token usage tracking
const tokenCounter = new Counter({
  name: 'cohere_tokens_total',
  help: 'Total tokens consumed',
  labelNames: ['endpoint', 'model', 'direction'], // direction: input|output
  registers: [registry],
});

// Error counter by type
const errorCounter = new Counter({
  name: 'cohere_errors_total',
  help: 'Cohere errors by status code',
  labelNames: ['endpoint', 'status_code'],
  registers: [registry],
});

// Rate limit headroom
const rateLimitGauge = new Gauge({
  name: 'cohere_rate_limit_remaining',
  help: 'Remaining rate limit capacity',
  labelNames: ['endpoint'],
  registers: [registry],
});

import { CohereClientV2, CohereError, CohereTimeoutError } from 'cohere-ai';

const cohere = new CohereClientV2();

async function instrumentedCall<T>(
  endpoint: string,
  model: string,
  operation: () => Promise<T>
): Promise<T> {
  const timer = requestDuration.startTimer({ endpoint, model });

  try {
    const result = await operation();
    requestCounter.inc({ endpoint, model, status: 'success' });
    timer();

    // Track tokens from response
    const usage = (result as any)?.usage?.billedUnits;
    if (usage) {
      if (usage.inputTokens) {
        tokenCounter.inc({ endpoint, model, direction: 'input' }, usage.inputTokens);
      }
      if (usage.outputTokens) {
        tokenCounter.inc({ endpoint, model, direction: 'output' }, usage.outputTokens);
      }
    }

    return result;
  } catch (err) {
    requestCounter.inc({ endpoint, model, status: 'error' });
    timer();

    if (err instanceof CohereError) {
      errorCounter.inc({ endp

// Use smaller models for latency-sensitive paths
function selectModel(latencyBudgetMs: number): string {
  if (latencyBudgetMs < 1000) return 'command-r7b-12-2024';   // 7B, fastest
  if (latencyBudgetMs < 3000) return 'command-r-08-2024';      // Mid-tier
  return 'command-a-03-2025';                                    // Best quality
}

// Pair with maxTokens to control output length
await cohere.chat({
  model: selectModel(1500),
  messages: [{ role: 'user', content: query }],
  maxTokens: 200,  // Shorter output = lower latency
});

// Non-streaming: user waits for entire response (800ms-5s)
// Streaming: first token arrives in ~200ms

async function streamForUI(message: string): Promise<string> {
  const stream = await cohere.chatStream({
    model: 'command-a-03-2025',
    messages: [{ role: 'user', content: message }],
  });

  let fullText = '';
  for await (const event of stream) {
    if (event.type === 'content-delta') {
      const text = event.delta?.message?.content?.text ?? '';
      fullText += text;
      // Emit to frontend immediately — perceived latency drops to ~200ms
    }
  }
  return fullText;
}

// BAD: 1000 texts = 1000 API calls
for (const text of texts) {
  await cohere.embed({ model: 'embed-v4.0', texts: [text], ... });
}

// GOOD: 1000 texts = 11 API calls (96 per batch)
async function batchEmbed(texts: string[]): Promise<number[][]> {
  const BATCH = 96;

// /api/health
import { CohereClientV2, CohereError } from 'cohere-ai';

const cohere = new CohereClientV2();

export async function GET() {
  const start = Date.now();
  let cohereStatus: 'healthy' | 'degraded' | 'down' = 'down';

  try {
    // Cheapest possible healt

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');
}

import PQueue from 'p

my-cohere-app/
├── src/
│   ├── cohere/
│   │   ├── client.ts           # CohereClientV2 singleton
│   │   ├── models.ts           # Model selection logic
│   │   ├── types.ts            # Cohere-specific types
│   │   └── errors.ts           # Error classification
│   ├── services/
│   │   ├── chat.ts             # Chat completions + streaming
│   │   ├── rag.ts              # RAG pipeline (embed → rerank → chat)
│   │   ├── agents.ts           # Tool-use agent loops
│   │   ├── embed.ts            # Batch embedding + caching
│   │   ├── rerank.ts           # Document reranking
│   │   └── classify.ts         # Few-shot classification
│   ├── tools/                  # Tool definitions for agents
│   │   ├── registry.ts         # Tool name → executor mapping
│   │   ├── search.ts
│   │   └── calculator.ts
│   ├── api/
│   │   ├── chat.ts             # POST /api/chat (streaming)
│   │   ├── embed.ts            # POST /api/embed
│   │   └── health.ts           # GET /api/health
│   └── cache/
│       └── embeddings.ts       # LRU cache for embeddings
├── tests/
│   ├── unit/
│   │   ├── chat.test.ts
│   │   ├── rag.test.ts
│   │   └── fixtures/           # Mocked API responses
│   └── integration/
│       └── cohere.test.ts      # Real API tests (gated)
├── config/
│   ├── models.json             # Model selection per environment
│   └── tools.json              # Tool definitions
└── package.json

┌─────────────────────────────────────────┐
│             API Layer                    │
│   (Express/Next.js routes, SSE stream)  │
├─────────────────────────────────────────┤
│           Service Layer                  │
│  (RAG pipeline, agent loop, classify)   │
├─────────────────────────────────────────┤
│          Cohere Layer                    │
│   (CohereClientV2, retry, model select) │
├─────────────────────────────────────────┤
│         Infrastructure Layer             │
│    (Embed cache, tool registry, queue)   │
└─────────────────────────────────────────┘

// src/cohere/client.ts
import { CohereClientV2, CohereError, CohereTimeoutError } from 'cohere-ai';

let instance: CohereClientV2 | null = null;

export function getCohere(): CohereClientV2 {
  if (!instance) {
    instance = new CohereClientV2({
      token: process.env.CO_API_KEY,
    });
  }
  return instance;
}

// src/cohere/models.ts
export const MODELS = {
  chat: {
    premium: 'command-a-03-2025',
    standard: 'command-r-08

// src/cohere/client.ts
import { CohereClientV2, CohereError, CohereTimeoutError } from 'cohere-ai';

let instance: CohereClientV2 | null = null;

export function getCohere(): CohereClientV2 {
  if (!instance) {
    if (!process.env.CO_API_KEY) {
      throw new Error('CO_API_KEY environment variable is required');
    }
    instance = new CohereClientV2({
      token: process.env.CO_API_KEY,
    });
  }
  return instance;
}

export async function withRetry<T>(
  operation: () => Promise<T>,
  maxRetries = 3,
  baseDelayMs = 1000
): Promise<T> {
  for (let attempt = 1; attempt <= maxRetries; attempt++) {
    try {
      return await operation();
    } catch (err) {
      if (attempt === maxRetries) throw err;

      // Only retry on rate limits (429) and server errors (5xx)
      if (err instanceof CohereError) {
        const status = err.statusCode;
        if (status && status !== 429 && status < 500) throw err;
      } else if (!(err instanceof CohereTimeoutError)) {
        throw err;
      }

      const delay = baseDelayMs * Math.pow(2, attempt - 1);
      const jitter = Math.random() * 500;
      await new Promise(r => setTimeout(r, delay + jitter));
    }
  }
  throw new Error('Unreachable');
}

// src/cohere/chat.ts
import { getCohere, withRetry } from './client';

interface ChatOptions {
  message: string;
  systemPrompt?: string;
  model?: string;
  maxTokens?: number;
  temperature?: number;
  documents?: Array<{ id?: string; data: Record<string, string> }>;
}

export async function chat(options: ChatOptions): Promise<string> {
  const cohere = getCohere();

  const messages: Array<{ role: string; content: string }> = [];

  if (options.systemPrompt) {
    messages.push({ role: 'system', content: options.systemPrompt });
  }
  messages.push({ role: 'user', content: options.message });

  const response = await withRetry(() =>
    cohere.chat({
      model: options.model ?? 'command-a-03-2025',
      messages,
      maxTokens: options.maxTokens,
      temperature: options.temperature,
      documents: options.documents,
    })
  );

  return response.message?.content?.[0]?.text ?? '';
}

                

                  
                    
                  
                  
                  cohere-security-basics
                  SKILL.md
                  View full skill →
                
                

                  
'Apply Cohere security best practices for API key management and access.
                  

                      ReadWriteGrep
                    
                
                

                  
Cohere Security Basics
Overview
Security best practices for Cohere API keys, request validation, and data protection. Cohere uses bearer token auth with trial and production key tiers.
Prerequisites

Cohere account at dashboard.cohere.com
Understanding of environment variables
Secret management solution for production

Instructions
Step 1: API Key Management

# NEVER hardcode keys — use environment variables
export CO_API_KEY="your-key-here"

# .env file (MUST be git-ignored)
CO_API_KEY=your-key-here

# .gitignore (mandatory entries)
.env
.env.local
.env.*.local

Key types:

Trial keys — free, rate-limited, for development only
Production keys — metered billing, for live applications

Step 2: Runtime Validation

import { CohereClientV2 } from 'cohere-ai';

function createSecureClient(): CohereClientV2 {
  const apiKey = process.env.CO_API_KEY;

  if (!apiKey) {
    throw new Error('CO_API_KEY is required. Set it as an environment variable.');
  }

  // Basic key format check
  if (apiKey.length < 20) {
    throw new Error('CO_API_KEY appears malformed. Check dashboard.cohere.com.');
  }

  return new CohereClientV2({ token: apiKey });
}

Step 3: Key Rotation Procedure

# 1. Generate new key in Cohere dashboard
#    → dashboard.cohere.com → API Keys → Create new key

# 2. Deploy new key (keep old key active)
# Vercel:
vercel env add CO_API_KEY production

# AWS:
aws secretsmanager update-secret --secret-id cohere/api-key --secret-string "new-key"

# GCP:
echo -n "new-key" | gcloud secrets versions add cohere-api-key --data-file=-

# 3. Verify new key works
curl -s -o /dev/null -w "%{http_code}" \
  -H "Authorization: Bearer NEW_KEY" \
  -H "Content-Type: application/json" \
  https://api.cohere.com/v2/chat \
  -d '{"model":"command-r7b-12-2024","messages":[{"role":"user","content":"test"}]}'
# Should return 200

# 4. Revoke old key in dashboard
# 5. Monitor for 401 errors after revocation

Step 4: Request Data Protection

// Scrub PII before sending to Cohere API
const PII_PATTERNS: [string, RegExp][] = [
  ['email', /[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}/g],
  ['phone', /\b\d{3}[-.]?\d{3}[-.]?\d{4}\b/g],
  ['ssn', /\b\d{3}-\d{2}-\d{4}\b/g],
];

function scrubPII(text: string): string {
  let scrubbed = text;
  for (const [type, regex] of PII_PATTERNS) {
    scrubbed = scrubbed.replace(regex, `[REDACTED_${type.toUpperCase()}]`);
  }
  return scrubbed;
}


                

              

# TypeScript
npm list cohere-ai

# Python
pip show cohere

# Latest versions
npm view cohere-ai version
pip index versions cohere

git checkout -b upgrade/cohere-v2-migration
npm install cohere-ai@latest
# or: pip install cohere --upgrade

// v1
import { CohereClient } from 'cohere-ai';
const cohere = new CohereClient({ token: '...' });

// v2
import { CohereClientV2 } from 'cohere-ai';
const cohere = new CohereClientV2({ token: '...' });

// v1 — single message string
const response = await cohere.chat({
  message: 'Hello',
  preamble: 'You are helpful.',
  chatHistory: [
    { role: 'USER', message: 'Hi' },
    { role: 'CHATBOT', message: 'Hello!' },
  ],
});
console.log(response.text);

// v2 — OpenAI-compatible messages array, model required
const response = await cohere.chat({
  model: 'command-a-03-2025',  // REQUIRED in v2
  messages: [
    { role: 'system', content: 'You are helpful.' },
    { role: 'user', content: 'Hi' },
    { role: 'assistant', content: 'Hello!' },
    { role: 'user', content: 'Hello' },
  ],
});
console.log(response.message?.content?.[0]?.text);

// v1 — model optional, no embedding_types
const response = await cohere.embed({
  texts: ['hello'],
});

// v2 — model, inputType, embeddingTypes all REQUIRED
const response = await cohere.embed({
  model: 'embed-v4.0',          // Required
  texts: ['hello'],
  inputType: 'search_document', // Required for v3+ models
  embeddingTypes: ['float'],    // Required for v3+ models
});

// v1
const respons

import { CohereClientV2 } from 'cohere-ai';

const cohere = new CohereClientV2();

async function handleStream(userMessage: string) {
  const stream = await cohere.chatStream({
    model: 'command-a-03-2025',
    messages: [{ role: 'user', content: userMessage }],
  });

  for await (const event of stream) {
    switch (event.type) {
      // Text content streaming
      case 'content-start':
        console.log('--- Generation started ---');
        break;

      case 'content-delta':
        const text = event.delta?.message?.content?.text ?? '';
        process.stdout.write(text);
        break;

      case 'content-end':
        console.log('\n--- Generation complete ---');
        break;

      // Citation events (when using documents)
      case 'citation-start':
        console.log('Citation:', event.delta?.message?.citations);
        break;

      // Tool call events (when using tools)
      case 'tool-call-start':
        console.log('Tool call started:', event.delta?.message?.toolCalls?.function?.name);
        break;

      case 'tool-call-delta':
        // Streaming tool arguments
        break;

      case 'tool-call-end':
        console.log('Tool call complete');
        break;

      // Message lifecycle
      case 'message-start':
        console.log('Message ID:', event.id);
        break;

      case 'message-end':
        console.log('Finish reason:', event.delta?.finishReason);
        console.log('Usage:', event.delta?.usage);
        break;
    }
  }
}

async function streamRAG(query: string, docs: string[]) {
  const stream = await cohere.chatStream({
    model: 'command-a-03-2025',
    messages: [{ role: 'user', content: query }],
    documents: docs.map((text, i) => ({
      id: `doc-${i}`,
      data: { text },
    })),
  });

  let fullText = '';
  const citations: Array<{ start: number; end: number; text: string; sources: string[] }> = [];

  for await (const event of stream) {
    if (event.type === 'content-delta') {
      const chunk = event.delta?.message?.content?.text ?? '';
      fullText