1// Model options
2- gpt-4o (latest, best)
3- gpt-4o-mini (fast, cheap)
4- gpt-4-turbo (powerful)
5- gpt-3.5-turbo (budget)
6
7// Parameters
8Temperature: 0.7  // 0-2, creativity
9Max Tokens: 1000  // Response length
10Top P: 1          // Nucleus sampling

1// Input
2{
3  "messages": [
4    { "role": "system", "content": "You are a helpful assistant" },
5    { "role": "user", "content": "Explain AI in simple terms" }
6  ]
7}
8
9// Output
10{
11  "content": "AI, or Artificial Intelligence, is like teaching computers to think...",
12  "model": "gpt-4o-mini",
13  "usage": { "total_tokens": 150 }
14}

1Model: claude-3-5-sonnet-20241022
2Max Tokens: 4096
3Temperature: 0.7
4System Prompt: "You are a Vietnamese customer support agent..."

1// Ollama node config
2Base URL: http://localhost:11434
3Model: llama2
4Keep Alive: 5m

1{
2  "prompt": "Write a product description for: {{ $json.product_name }}",
3  "product_name": "Vietnamese Coffee Filter Set"
4}

1// Messages configuration
2[
3  {
4    "role": "system",
5    "content": "You are a marketing copywriter specializing in Vietnamese products."
6  },
7  {
8    "role": "user", 
9    "content": "{{ $json.prompt }}"
10  }
11]

1// Dynamic prompts với n8n expressions
2System: "You are an assistant for {{ $env.COMPANY_NAME }}"
3User: "Help customer {{ $json.customer_name }} with: {{ $json.question }}"

1// Code node for complex prompts
2const customerData = $input.first().json;
3
4const prompt = `
5You are a customer support agent.
6
7Customer Information:
8- Name: ${customerData.name}
9- Account Type: ${customerData.accountType}
10- Previous Issues: ${customerData.issueCount}
11
12Current Request:
13${customerData.message}
14
15Guidelines:
16- Be polite and professional
17- Refer to customer by name
18- If VIP customer (Premium account), offer priority support
19`;
20
21return [{ json: { prompt } }];

1// OpenAI output structure
2{
3  "message": {
4    "content": "Generated text here...",
5    "role": "assistant"
6  },
7  "usage": {
8    "prompt_tokens": 50,
9    "completion_tokens": 100,
10    "total_tokens": 150
11  }
12}
13
14// Access in next node
15{{ $json.message.content }}

1// Request JSON output
2System: "Always respond in JSON format with keys: summary, sentiment, action_items"
3
4// Parse response
5const response = $input.first().json.message.content;
6
7try {
8  const parsed = JSON.parse(response);
9  return [{ json: parsed }];
10} catch (e) {
11  return [{ json: { error: "Failed to parse JSON", raw: response } }];
12}

1// After OpenAI node
2const usage = $input.first().json.usage;
3
4// Calculate cost (GPT-4o-mini pricing)
5const inputCost = (usage.prompt_tokens / 1000) * 0.00015;
6const outputCost = (usage.completion_tokens / 1000) * 0.0006;
7const totalCost = inputCost + outputCost;
8
9// Log to database or tracking service
10return [{
11  json: {
12    ...input.first().json,
13    cost: {
14      input_tokens: usage.prompt_tokens,
15      output_tokens: usage.completion_tokens,
16      total_cost_usd: totalCost.toFixed(6)
17    }
18  }
19}];

1// Strategy 1: Use smaller models for simple tasks
2const task = $json.task_type;
3const model = task === "simple_classification" 
4  ? "gpt-3.5-turbo" 
5  : "gpt-4o-mini";
6
7// Strategy 2: Limit max tokens
8Max Tokens: 500  // Don't allow long responses if not needed
9
10// Strategy 3: Cache similar requests
11// Check Redis/database for cached response before calling API

1// Wrap AI calls in error handling
2try {
3  const result = await openAICall();
4  return [{ json: { success: true, result } }];
5} catch (error) {
6  if (error.code === 'rate_limit_exceeded') {
7    // Wait and retry
8    await sleep(5000);
9    return [{ json: { retry: true } }];
10  }
11  
12  if (error.code === 'context_length_exceeded') {
13    // Truncate input
14    return [{ json: { error: "Input too long", suggestion: "Shorten prompt" } }];
15  }
16  
17  return [{ json: { error: error.message } }];
18}