Imported from jeremylongshore/tons-of-skills-marketplace (plugins/saas-packs/cohere-pack/skills/cohere-sdk-patterns/SKILL.md). Install upstream with npx skills add jeremylongshore/tons-of-skills-marketplace --skill cohere-sdk-patterns. Copyright stays with the author (MIT).

Cohere SDK Patterns

Overview

Production-ready patterns for the cohere-ai TypeScript SDK (CohereClientV2) and Python cohere package. Real model names, real API shapes, real error types.

Prerequisites

• cohere-ai v7+ installed (TypeScript) or cohere v5+ (Python)
• Familiarity with async/await patterns
• Understanding of Cohere API v2 endpoints

Instructions

Pattern 1: Singleton Client with Retry

// 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');
}

Pattern 2: Type-Safe Chat Wrapper

// 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 ?? '';
}

Pattern 3: Streaming Chat

// src/cohere/stream.ts
export async function* streamChat(
  message: string,
  model = 'command-a-03-2025'
): AsyncGenerator<string> {
  const cohere = getCohere();

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

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

// Usage
for await (const chunk of streamChat('Explain RAG in 3 sentences')) {
  process.stdout.write(chunk);
}

Pattern 4: Batch Embedding

// src/cohere/embed.ts
type InputType = 'search_document' | 'search_query' | 'classification' | 'clustering';

export async function embedTexts(
  texts: string[],
  inputType: InputType = 'search_document',
  model = 'embed-v4.0'
): Promise<number[][]> {
  const cohere = getCohere();

  // Cohere embed accepts up to 96 texts per call
  const BATCH_SIZE = 96;
  const allEmbeddings: number[][] = [];

  for (let i = 0; i < texts.length; i += BATCH_SIZE) {
    const batch = texts.slice(i, i + BATCH_SIZE);
    const response = await withRetry(() =>
      cohere.embed({
        model,
        texts: batch,
        inputType,
        embeddingTypes: ['float'],
      })
    );
    allEmbeddings.push(...response.embeddings.float);
  }

  return allEmbeddings;
}

Pattern 5: Rerank with Type Safety

// src/cohere/rerank.ts
interface RerankResult {
  text: string;
  score: number;
  originalIndex: number;
}

export async function rerankDocuments(
  query: string,
  documents: string[],
  topN = 5,
  model = 'rerank-v3.5'
): Promise<RerankResult[]> {
  const cohere = getCohere();

  const response = await withRetry(() =>
    cohere.rerank({ model, query, documents, topN })
  );

  return response.results.map(r => ({
    text: documents[r.index],
    score: r.relevanceScore,
    originalIndex: r.index,
  }));
}

Pattern 6: Structured JSON Output

export async function chatJSON<T>(
  message: string,
  schema?: Record<string, unknown>
): Promise<T> {
  const cohere = getCohere();

  const response = await cohere.chat({
    model: 'command-a-03-2025',
    messages: [{ role: 'user', content: `${message}\n\nRespond in valid JSON.` }],
    responseFormat: schema
      ? { type: 'json_object', jsonSchema: schema }
      : { type: 'json_object' },
  });

  const text = response.message?.content?.[0]?.text ?? '{}';
  return JSON.parse(text) as T;
}

Python Equivalents

import cohere
from cohere import ClientV2

# Singleton
_client: ClientV2 | None = None

def get_cohere() -> ClientV2:
    global _client
    if _client is None:
        _client = ClientV2()  # reads CO_API_KEY
    return _client

# Chat
def chat(message: str, model: str = "command-a-03-2025") -> str:
    co = get_cohere()
    response = co.chat(
        model=model,
        messages=[{"role": "user", "content": message}],
    )
    return response.message.content[0].text

# Embed
def embed(texts: list[str], input_type: str = "search_document") -> list[list[float]]:
    co = get_cohere()
    response = co.embed(
        model="embed-v4.0",
        texts=texts,
        input_type=input_type,
        embedding_types=["float"],
    )
    return response.embeddings.float

Error Handling

Output

Return validated domain results or classified retryable/terminal errors with safe correlation data, model/version, and operation context. The client boundary must not log API keys, complete prompts, retrieved documents, embeddings, or personal data beyond the caller’s approved need.

Examples

Use a staging client with a scoped key to embed a synthetic document, validate the expected embedding type and dimensions, then assert a malformed response becomes a contract error without caching it. On a 429, reschedule the same bounded job; on 401, stop and repair the secret reference through its owner.

Resources

• Cohere TypeScript SDK
• Cohere Python SDK
• API v2 Reference

Next Steps

Apply patterns in cohere-core-workflow-a for RAG workflows.