Workers AI

Run AI inference at the edge using Workers AI and industry-standard SDKs like OpenAI. Deploy LLM-powered applications with structured outputs, streaming responses, and AI Gateway integration.

FIRST: Installation

npm install openai

Optional dependencies for advanced use cases:

npm install ai @ai-sdk/openai  # For streaming with Vercel AI SDK

When to Use

Quick Reference

Structured JSON Outputs

Workers AI supports structured JSON outputs using the OpenAI SDK's response_format API. This ensures the model returns data matching your schema.

import { OpenAI } from "openai";

interface Env {
  OPENAI_API_KEY: string;
}

// Define your JSON schema
const CalendarEventSchema = {
  type: 'object',
  properties: {
    name: { type: 'string' },
    date: { type: 'string' },
    participants: { type: 'array', items: { type: 'string' } },
  },
  required: ['name', 'date', 'participants']
};

export default {
  async fetch(request: Request, env: Env) {
    const client = new OpenAI({
      apiKey: env.OPENAI_API_KEY,
    });

    const response = await client.chat.completions.create({
      model: 'gpt-4o-2024-08-06',
      messages: [
        { role: 'system', content: 'Extract the event information.' },
        { role: 'user', content: 'Alice and Bob are going to a science fair on Friday.' },
      ],
      // Request structured JSON output with schema validation
      response_format: {
        type: 'json_schema',
        schema: CalendarEventSchema,
      },
    });

    // Parsed according to your schema
    const event = response.choices[0].message.parsed;

    return Response.json({
      calendar_event: event,
    });
  }
}

wrangler.jsonc:

{
  "name": "my-ai-app",
  "main": "src/index.ts",
  "compatibility_date": "2025-01-17",
  "observability": {
    "enabled": true
  }
}

Streaming Responses

For real-time chat experiences, use streaming to send tokens as they're generated.

import { OpenAI } from "openai";

interface Env {
  OPENAI_API_KEY: string;
}

export default {
  async fetch(request: Request, env: Env) {
    const client = new OpenAI({
      apiKey: env.OPENAI_API_KEY,
    });

    const stream = await client.chat.completions.create({
      model: 'gpt-4o',
      messages: [
        { role: 'user', content: 'Tell me a story about the edge.' }
      ],
      stream: true,
    });

    // Create a ReadableStream for SSE
    const encoder = new TextEncoder();
    const readable = new ReadableStream({
      async start(controller) {
        try {
          for await (const chunk of stream) {
            const content = chunk.choices[0]?.delta?.content;
            if (content) {
              controller.enqueue(encoder.encode(`data: ${JSON.stringify({ content })}\n\n`));
            }
          }
          controller.enqueue(encoder.encode('data: [DONE]\n\n'));
          controller.close();
        } catch (error) {
          controller.error(error);
        }
      },
    });

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

AI Gateway Integration

AI Gateway provides caching, rate limiting, analytics, and request logging for your AI requests. Configure it by setting the baseUrl in your OpenAI client.

import { OpenAI } from "openai";

interface Env {
  OPENAI_API_KEY: string;
  AI_GATEWAY_ACCOUNT_ID: string;
  AI_GATEWAY_ID: string;
}

export default {
  async fetch(request: Request, env: Env) {
    const client = new OpenAI({
      apiKey: env.OPENAI_API_KEY,
      // Route requests through AI Gateway
      baseUrl: `https://gateway.ai.cloudflare.com/v1/${env.AI_GATEWAY_ACCOUNT_ID}/${env.AI_GATEWAY_ID}/openai`
    });

    const response = await client.chat.completions.create({
      model: 'gpt-4o',
      messages: [
        { role: 'user', content: 'Hello, world!' }
      ],
    });

    return Response.json(response.choices[0].message);
  }
}

Benefits of AI Gateway:

• Caching: Reduce costs by caching identical requests
• Rate Limiting: Protect against abuse and control costs
• Analytics: Monitor token usage, latency, and error rates
• Logging: Inspect requests and responses for debugging
• Multi-provider: Works with OpenAI, Anthropic, Azure, and more

Model Selection

Choose models based on your use case:

Choosing the right model:

• Structured extraction: Use GPT-4o with json_schema
• Chat interfaces: Use GPT-4o or Claude 3.5 Sonnet with streaming
• High volume/low latency: Use GPT-4o-mini or Claude 3 Haiku
• Complex reasoning: Use GPT-4o or Claude 3.5 Sonnet

Response Formats

Workers AI supports multiple response format options:

// Option 1: JSON Schema (recommended for structured extraction)
response_format: {
  type: 'json_schema',
  schema: {
    type: 'object',
    properties: {
      name: { type: 'string' },
      age: { type: 'number' },
    },
    required: ['name']
  }
}

// Option 2: JSON Object (parse manually)
response_format: {
  type: 'json_object'
}
// Remember to prompt the model to return JSON

// Option 3: Text (default)
// No response_format specified - returns plain text

Generating Embeddings

Use embeddings for semantic search, RAG, and similarity matching. Combine with Vectorize for storage.

import { OpenAI } from "openai";

interface Env {
  OPENAI_API_KEY: string;
  VECTORIZE: VectorizeIndex;
}

export default {
  async fetch(request: Request, env: Env) {
    const client = new OpenAI({
      apiKey: env.OPENAI_API_KEY,
    });

    const text = "Cloudflare Workers run at the edge";

    // Generate embedding
    const response = await client.embeddings.create({
      model: 'text-embedding-3-small',
      input: text,
    });

    const vector = response.data[0].embedding;

    // Store in Vectorize
    await env.VECTORIZE.upsert([
      {
        id: '1',
        values: vector,
        metadata: { text }
      }
    ]);

    return Response.json({ 
      dimensions: vector.length,
      stored: true 
    });
  }
}

wrangler.jsonc with Vectorize binding:

{
  "vectorize": [
    {
      "binding": "VECTORIZE",
      "index_name": "my-embeddings-index"
    }
  ]
}

Error Handling

Always handle AI API errors gracefully:

export default {
  async fetch(request: Request, env: Env) {
    const client = new OpenAI({
      apiKey: env.OPENAI_API_KEY,
    });

    try {
      const response = await client.chat.completions.create({
        model: 'gpt-4o',
        messages: [{ role: 'user', content: 'Hello!' }],
      });

      return Response.json(response.choices[0].message);
    } catch (error) {
      // Handle rate limits
      if (error.status === 429) {
        return Response.json(
          { error: 'Rate limit exceeded. Please try again later.' },
          { status: 429 }
        );
      }

      // Handle invalid requests
      if (error.status === 400) {
        return Response.json(
          { error: 'Invalid request. Check your parameters.' },
          { status: 400 }
        );
      }

      // Generic error
      console.error('AI request failed:', error);
      return Response.json(
        { error: 'Internal server error' },
        { status: 500 }
      );
    }
  }
}

Detailed References

• references/models.md - Model capabilities, pricing, and selection guide
• references/streaming.md - Streaming patterns, SSE, and client integration
• references/dynamic-model-discovery.md - Programmatically discover models and capabilities at runtime
• references/testing.md - Mocking AI responses, remote bindings, testing different models

Best Practices

1. Use structured outputs: Set response_format with json_schema for reliable data extraction
2. Enable observability: Set observability.enabled: true in wrangler.jsonc
3. Stream for chat: Use streaming responses for better user experience
4. Cache with AI Gateway: Route requests through AI Gateway to cache and monitor
5. Handle errors: Always catch and handle API errors gracefully
6. Choose right model: Balance cost, speed, and capability based on your use case
7. Validate inputs: Sanitize user inputs before sending to AI models
8. Set timeouts: Use appropriate timeouts for long-running requests
9. Use embeddings wisely: Batch embedding generation when possible
10. Monitor token usage: Track costs through AI Gateway analytics

Integration Patterns

Pattern 1: Chat with Message History

interface Env {
  OPENAI_API_KEY: string;
  KV: KVNamespace;
}

export default {
  async fetch(request: Request, env: Env) {
    const { userId, message } = await request.json();
    const client = new OpenAI({ apiKey: env.OPENAI_API_KEY });

    // Get message history from KV
    const historyJson = await env.KV.get(`chat:${userId}`);
    const history = historyJson ? JSON.parse(historyJson) : [];

    // Add user message
    history.push({ role: 'user', content: message });

    // Get AI response
    const response = await client.chat.completions.create({
      model: 'gpt-4o',
      messages: history,
    });

    const assistantMessage = response.choices[0].message;
    history.push(assistantMessage);

    // Store updated history
    await env.KV.put(`chat:${userId}`, JSON.stringify(history), {
      expirationTtl: 3600 // 1 hour
    });

    return Response.json({ message: assistantMessage.content });
  }
}

Pattern 2: RAG with Vectorize

interface Env {
  OPENAI_API_KEY: string;
  VECTORIZE: VectorizeIndex;
}

export default {
  async fetch(request: Request, env: Env) {
    const { query } = await request.json();
    const client = new OpenAI({ apiKey: env.OPENAI_API_KEY });

    // Generate query embedding
    const embeddingResponse = await client.embeddings.create({
      model: 'text-embedding-3-small',
      input: query,
    });

    // Search similar documents
    const results = await env.VECTORIZE.query(embeddingResponse.data[0].embedding, {
      topK: 3,
    });

    // Build context from results
    const context = results.matches
      .map(match => match.metadata.text)
      .join('\n\n');

    // Generate answer with context
    const response = await client.chat.completions.create({
      model: 'gpt-4o',
      messages: [
        { 
          role: 'system', 
          content: `Answer questions using this context:\n\n${context}` 
        },
        { role: 'user', content: query }
      ],
    });

    return Response.json({
      answer: response.choices[0].message.content,
      sources: results.matches.map(m => m.metadata),
    });
  }
}

Common Pitfalls

1. Not handling rate limits: Always catch 429 errors and implement backoff
2. Ignoring token limits: Monitor and truncate input to stay within model limits
3. Not caching: Use AI Gateway or KV to cache responses for identical requests
4. Blocking on responses: Use streaming for better perceived performance
5. Missing error boundaries: Wrap AI calls in try-catch blocks
6. Hardcoding API keys: Always use environment bindings
7. Not validating schemas: Test your JSON schemas thoroughly
8. Overfitting prompts: Keep system prompts concise and clear