cf-workers-ai

Workers AI

AI Call Pattern

const result = await env.AI.run(
  '@cf/mistralai/mistral-small-3.1-24b-instruct',
  {
    messages: [
      { role: 'system', content: 'You are a helpful assistant.' },
      { role: 'user', content: prompt }
    ],
    max_tokens: 500
  }
);

Key Concepts

• AI binding: Declared in wrangler.toml, accessed via env.AI
• Model invocation: env.AI.run(model, options)
• Available models: Mistral, Llama, and others on Cloudflare's edge
• Prompt engineering: System + user message pattern
• Token management: Set max_tokens, truncate long inputs
• Streaming: Use stream: true for streaming responses

Anti-Patterns

• Ignoring token limits with AI models
• Not handling model errors/timeouts
• Sending unbounded user input without truncation

Reference

For detailed patterns, see references/workers-ai.md.