Using AI Nodes

Leverage LLMs and AI capabilities in your workflows.

LLM Node

The most commonly used AI node for text generation.

Basic Configuration

{
  provider: "openai",
  model: "gpt-4.1",
  systemPrompt: "You are a helpful assistant",
  userMessage: "{{trigger.body.question}}",
  temperature: 0.7,
  maxTokens: 500
}

Supported Providers

Provider	Models	Best For
OpenAI	gpt-4.1, gpt-4o, gpt-4o-mini, o3-mini	General purpose, function calling
Anthropic	claude-opus-4-5, claude-sonnet-4-5, claude-haiku-4-5	Instruction following, safety
Google	gemini-2.0-flash, gemini-2.0-pro	Multi-modal, large context
Cohere	command-r-plus, command-r	Retrieval, enterprise
x.ai	grok-3, grok-3-mini	Reasoning, analysis
Hugging Face	Various	Open source, specialized

Options

Option	Description	Default
Temperature	Creativity (0 = deterministic, 2 = very creative)	0.7
Max Tokens	Response length limit	4096
JSON Mode	Force structured JSON output	false
Stop Sequences	Tokens that stop generation	[]

Example: Content Summarizer

{
  provider: "anthropic",
  model: "claude-sonnet-4-5",
  systemPrompt: `You are a content summarizer.
    Summarize the provided text in 3 bullet points.
    Be concise but capture key information.`,
  userMessage: "{{http_request.output.body}}",
  temperature: 0.3
}

---

Extended Thinking

Enable deep reasoning for complex tasks with models that support extended thinking.

Supported Models

Model	Provider	Thinking Capability
claude-opus-4-5	Anthropic	Extended thinking
claude-sonnet-4-5	Anthropic	Extended thinking
o3	OpenAI	Extended thinking
o3-mini	OpenAI	Reasoning mode
grok-3	x.ai	Extended thinking

Configuration

{
  provider: "anthropic",
  model: "claude-sonnet-4-5",
  extendedThinking: {
    enabled: true,
    budgetTokens: 10000,  // Max tokens for thinking
    showThinking: false    // Include thinking in output
  },
  userMessage: "Analyze this complex problem: {{input.problem}}"
}

When to Use

Extended thinking is ideal for:

• Complex reasoning — Multi-step logic problems
• Code analysis — Understanding large codebases
• Research synthesis — Combining multiple sources
• Strategic planning — Business decisions

Example: Code Review

{
  provider: "openai",
  model: "o3",
  extendedThinking: { enabled: true, budgetTokens: 8000 },
  systemPrompt: `You are a senior code reviewer.
    Analyze the code for bugs, security issues, and improvements.
    Think through each issue carefully.`,
  userMessage: "Review this code:\n{{input.code}}"
}

---

Vision Node

Analyze images with AI.

Basic Configuration

{
  provider: "openai",
  model: "gpt-4o",
  image: "{{trigger.body.imageUrl}}",
  prompt: "Describe what you see in this image"
}

Image Sources

// URL
{
    image: "https://example.com/image.jpg";
}

// Base64
{
    image: "{{screenshot.output.base64}}";
}

// File upload
{
    image: "{{trigger.files[0].url}}";
}

Multiple Images

{
  model: "gpt-4o",
  images: [
    "{{input.image1}}",
    "{{input.image2}}"
  ],
  prompt: "Compare these two images and describe the differences"
}

Example: Product Analysis

{
  provider: "anthropic",
  model: "claude-sonnet-4-5",
  image: "{{trigger.body.productImage}}",
  prompt: `Analyze this product image:
    1. Identify the product type
    2. List visible features
    3. Estimate quality level
    4. Suggest improvements`,
  outputVariable: "product_analysis"
}

---

Audio Nodes

Speech-to-Text (STT)

Convert audio to text:

{
  type: "audio_input",
  provider: "openai",
  model: "whisper-1",
  audio: "{{trigger.files[0].url}}",
  language: "en",
  outputVariable: "transcription"
}

Supported formats: mp3, mp4, mpeg, mpga, m4a, wav, webm

Options:

Option	Description
language	ISO language code (auto-detect if omitted)
timestamps	Include word-level timestamps
diarization	Identify different speakers

Text-to-Speech (TTS)

Convert text to audio:

{
  type: "audio_output",
  provider: "openai",
  model: "tts-1-hd",
  text: "{{llm.output.content}}",
  voice: "alloy",
  outputFormat: "mp3",
  outputVariable: "audio_file"
}

Available voices: alloy, echo, fable, onyx, nova, shimmer

Example: Voice Assistant Pipeline

[Audio Input] → [STT] → [LLM] → [TTS] → [Audio Output]

// STT: Transcribe user speech
{
  type: "audio_input",
  audio: "{{trigger.audio}}",
  outputVariable: "user_speech"
}

// LLM: Generate response
{
  type: "llm",
  userMessage: "{{user_speech.text}}",
  outputVariable: "response"
}

// TTS: Convert to speech
{
  type: "audio_output",
  text: "{{response.content}}",
  voice: "nova",
  outputVariable: "reply_audio"
}

---

Router Node

Intelligently route workflows based on content analysis.

Configuration

{
  type: "router",
  input: "{{trigger.body.message}}",
  routes: [
    {
      name: "billing",
      description: "Questions about billing, payments, invoices",
      destination: "billing_workflow"
    },
    {
      name: "technical",
      description: "Technical support, bugs, errors",
      destination: "tech_support_workflow"
    },
    {
      name: "sales",
      description: "Pricing, plans, upgrades",
      destination: "sales_workflow"
    }
  ],
  defaultRoute: "general_support",
  confidenceThreshold: 0.7
}

How It Works

1. Router analyzes the input text
2. Classifies intent using LLM
3. Routes to matching destination
4. Falls back to default if confidence < threshold

Example: Support Ticket Router

{
  type: "router",
  input: "{{trigger.body.subject}}: {{trigger.body.description}}",
  model: "gpt-4o-mini",  // Faster model for classification
  routes: [
    {
      name: "urgent",
      description: "Service outages, critical bugs, security issues",
      destination: "urgent_escalation",
      priority: 1
    },
    {
      name: "bug_report",
      description: "Bug reports, error messages, unexpected behavior",
      destination: "engineering_queue"
    },
    {
      name: "feature_request",
      description: "New feature ideas, suggestions, improvements",
      destination: "product_backlog"
    }
  ],
  outputVariable: "routing_decision"
}

Accessing Routing Results

{
    {
        routing_decision.route;
    }
} // Selected route name
{
    {
        routing_decision.confidence;
    }
} // Confidence score (0-1)
{
    {
        routing_decision.reasoning;
    }
} // Why this route was chosen

---

Embeddings Node

Generate vector embeddings for semantic search.

{
  type: "embeddings",
  provider: "openai",
  model: "text-embedding-3-small",
  input: "{{trigger.body.text}}",
  outputVariable: "embedding"
}

Use for:

• Storing content in vector databases
• Semantic similarity search
• Document clustering

---

Best Practices

Model Selection

Task	Recommended Model
Simple classification	gpt-4o-mini, claude-haiku-4-5
Complex reasoning	claude-opus-4-5, o3
Code generation	claude-sonnet-4-5, gpt-4.1
Vision tasks	gpt-4o, claude-sonnet-4-5
Fast responses	gpt-4o-mini, claude-haiku-4-5

Prompt Optimization

• Be specific about output format
• Use examples for consistency
• Set clear boundaries
• Test with varied inputs

Error Handling

Add retry logic for rate limits:

{
  retryPolicy: {
    maxRetries: 3,
    retryableErrors: ["RATE_LIMIT", "TIMEOUT"],
    backoffMs: 1000
  }
}

Cost Management

• Use smaller models when possible
• Set max_tokens appropriately
• Cache common responses
• Monitor usage in dashboard

---

Next Steps

• Node Types Reference — All AI node details
• Error Handling — Robust workflows
• Building RAG Agent — Knowledge-powered AI