AI Nodes

AI nodes add artificial intelligence to your workflows: text generation with LLMs, knowledge base search, Retrieval-Augmented Generation, reusable prompts, and autonomous agents that orchestrate multiple tools.

This page is the single reference for every AI node — its parameters and outputs — combined with practical prompt-engineering guidance, common patterns, and troubleshooting.

Time to work through the guidance sections: 20-25 minutes

---

AI Node Overview

Node	Purpose	Use Case
LLM Integration	Generate text with AI models	Content creation, analysis, classification
Prompt Template	Create structured, reusable prompts	Consistent AI interactions
Core Agent	Multi-tool AI orchestrator	Complex, multi-step task automation
KB Search	Search knowledge bases	Find relevant documents
KB RAG Query	Question-answering from your data	AI answers grounded in documents
Text Response	Format AI outputs	User-facing responses

Node Selection Guide

Need	Use
Generate text	LLM Integration
Find documents	KB Search
Q&A from docs	KB RAG Query
Multi-tool AI tasks	Core Agent
Structured prompts	Prompt Template
Format / send a response	Text Response

---

LLM Integration Node

Generate text using AI language models configured through the integration system.

When to Use

• Generate content (emails, summaries, responses)
• Analyze text (sentiment, classification, extraction)
• Transform data (reformat, translate, simplify)
• Answer questions based on context

Configuration

Field	Type	Required	Description	Default
LLM Provider	Select	Yes	LLM integration to use (OpenAI, Claude, etc.)	—
Model	Select	Yes	Specific model to use	gpt-4o-mini
Temperature	Slider	No	Creativity level (0.0-2.0)	0.2
Timeout	Number	No	Request timeout in seconds (10-600)	—
Max Retries	Number	No	Retry attempts on failure (0-5)	—
System Prompt	Text	No	AI persona / instructions	"You are a helpful assistant"
Prompt Template	Text	Yes	User prompt with variables	{{message}}
Max Tokens	Number	No	Maximum response length	1000
JSON Mode	Toggle	No	Force JSON output format	false

Model Override

Use model_override for dynamic model selection at runtime:

Model Override: {{evaluator.recommended_model}}

Available Models

Models depend on your configured integrations. Common options:

OpenAI:

Model	Best For	Cost
gpt-4o	Complex reasoning	$$$
gpt-4o-mini	General use (recommended)	$
gpt-4-turbo	Fast + capable	$$
gpt-3.5-turbo	Simple tasks / budget	$

Anthropic Claude:

Model	Best For	Cost
claude-3-opus	Most capable	$$$
claude-3-sonnet	Balanced	$$
claude-3-haiku	Fast + cheap	$

Google:

Model	Best For	Cost
gemini-1.5-pro	Full featured	$$
gemini-1.5-flash	Quick responses	$

Prompt Template Syntax

Use variables in your prompts:

System Prompt:
You are a professional email writer for {{company_name}}.

Prompt Template:
Write a follow-up email to {{contact.first_name}} about {{deal.name}}.

Key points to include:
- Last meeting date: {{last_meeting.date}}
- Action items: {{last_meeting.notes}}
- Next steps: {{deal.next_steps}}

Keep the tone professional but friendly.

Output Variables

{{llmintegration_1.output}}            - Generated text
{{llmintegration_1.success}}           - true/false
{{llmintegration_1.model}}             - Model used
{{llmintegration_1.tokens_used}}       - Total tokens
{{llmintegration_1.prompt_tokens}}     - Input tokens
{{llmintegration_1.completion_tokens}} - Output tokens
{{llmintegration_1.cost}}              - Estimated cost

Best Practices

✅ Be specific in prompts
✅ Use system prompt for persona
✅ Include examples in complex tasks
✅ Set appropriate temperature
✅ Handle empty outputs

❌ Don't use vague instructions
❌ Don't skip system prompt
❌ Don't set max_tokens too low

---

Prompt Template Node

Create reusable, structured prompts with variable substitution.

When to Use

• Standardize AI interactions
• Separate prompt logic from LLM calls
• Create complex multi-part prompts
• Use prompts from your prompt library

Configuration

Field	Type	Required	Description
Template Type	Select	Yes	custom or prompt_library
Prompt	Text	Yes	Main prompt template
System Prompt	Text	No	Context / persona
Include System Prompt	Toggle	No	Add system prompt to output
Formatting	Select	No	markdown, plain, json, html
Selected Prompt IDs	Array	No	IDs from the Prompt Library

Variable Syntax

Basic variables:

Hello {{first_name}}, welcome to {{company_name}}!

Nested paths:

Order #{{order.id}} for {{order.customer.name}}

Array access:

First item: {{items[0].name}}
Last item: {{items[-1].name}}

Fallback values:

Hello {{first_name|"Valued Customer"}}

Type conversions:

Score: {{lead_score|int}}
Price: {{amount|float}}
Active: {{is_active|bool}}

Special variables:

Date: {{date}}  // Current date (YYYY-MM-DD)
Time: {{time}}  // Current time (HH:MM:SS)

Example Template

You are analyzing customer feedback.

Customer: {{contact.first_name}} {{contact.last_name}}
Company: {{contact.company.name}}
Message: {{message}}

Provide:
1. Sentiment analysis
2. Key topics mentioned
3. Recommended action

Output Variables

{{prompttemplate_1.output}}          - Processed prompt
{{prompttemplate_1.system_prompt}}   - System prompt
{{prompttemplate_1.variables_used}}  - Variables resolved

---

Core Agent Node

AI orchestrator that can use tools to complete complex, multi-step tasks autonomously.

When to Use

• Multi-step reasoning required
• Need to use multiple tools
• Dynamic decision-making
• Complex problem-solving

Configuration

Field	Type	Required	Description	Default
Name	Text	No	Display name for agent instance	—
LLM Provider	Select	Yes	AI provider integration	—
Model	Select	Yes	Model for reasoning	gpt-4o-mini
Temperature	Slider	No	Response creativity (0-1)	0.7
System Prompt	Text	Yes	Agent instructions and persona	"You are a helpful assistant"
Max Iterations	Number	No	Maximum tool-use cycles	10
Available Tools	Multi-select	No	Tools the agent can access	—
Memory Type	Select	No	none, buffer, summary	buffer
Max Memory Messages	Number	No	Messages to remember	50

How It Works

The agent follows a Think → Act → Observe loop:

1. Think: Analyze the task and decide what to do
2. Act: Use a tool to get information or take action
3. Observe: Process the tool's result
4. Repeat: Continue until the task is complete or max iterations is reached

Available Tools

The agent can access any workflow node configured as a tool:

• Knowledge Base Search
• CRM Manager (contacts, companies, deals)
• Database Query
• API Request
• Email Send
• And more...

System Prompt Example

You are a customer support agent for TechCorp.

Your capabilities:
- Search the knowledge base for product information
- Look up customer orders in the CRM
- Send follow-up emails

Guidelines:
- Always search the KB before answering product questions
- Be polite and professional
- If you can't find an answer, say so honestly
- Summarize findings clearly

Output Variables

{{coreagent_1.output}}                   - Final response
{{coreagent_1.success}}                  - Completion status
{{coreagent_1.steps}}                    - Reasoning steps taken
{{coreagent_1.tool_calls}}               - Tools used
{{coreagent_1.metrics.total_duration}}   - Time taken
{{coreagent_1.metrics.llm_calls}}        - LLM calls made
{{coreagent_1.metrics.tokens_used}}      - Total tokens

tip

For chat-style, autonomous experiences with skills, see Agentic Mode.

---

KB Search Node

Search your knowledge bases for relevant documents using semantic search.

When to Use

• Find documents matching a query
• Provide context to AI prompts
• Search product documentation
• Look up policies and procedures

Configuration

Field	Type	Required	Description	Default
Vector DB Integration	Select	Conditional	Vector database	—
Knowledge Base	Select	Conditional	KB to search (or use kb_id)	—
Query	Text	Yes	Search query	—
Top K	Number	No	Candidates to retrieve	20
Limit	Number	No	Results to return	5
Threshold	Number	No	Minimum similarity (0-1)	—
Use Hybrid Search	Toggle	No	BM25 + Vector search	true
BM25 Weight	Slider	No	Keyword vs semantic balance (0-1)	0.5
Metadata Filters	Object	No	Filter by metadata fields	—

Search Modes

Semantic Search (default):

• Finds conceptually similar content
• Works even with different wording
• Good for: "What's our refund policy?"

Query: "customer onboarding process"

Hybrid Search (recommended):

• Combines keyword + semantic
• Better for specific terms
• Good for: "Order #12345 status"

Use Hybrid Search: true
BM25 Weight: 0.5
Query: "onboarding checklist"

With variables:

query: "{{contact.inquiry}} for {{contact.product}}"

Metadata filtering:

Query: "return policy"
Metadata Filters: {
  "document_type": "policy",
  "department": "support",
  "entity_type": "deals",
  "status": "active"
}

Output Variables

{{kbsearch_1.results}}               - Array of matches
{{kbsearch_1.results[0].content}}    - First result content
{{kbsearch_1.results[0].similarity}} - Match score (0-1)
{{kbsearch_1.results[0].metadata}}   - Document metadata
{{kbsearch_1.total_count}}           - Total matches found
{{kbsearch_1.success}}               - true/false

info

Learn how to build and populate knowledge bases in Knowledge Bases.

---

KB RAG Query Node

Ask questions and get AI-generated answers from your knowledge base using Retrieval-Augmented Generation.

When to Use

• Answer questions from documentation
• Customer support automation
• Internal Q&A systems
• Policy inquiries

Configuration

Field	Type	Required	Description	Default
Knowledge Base	Select	Yes	KB to query	—
Question	Text	Yes	Question to answer	—
Model	Select	No	AI model for answers	gpt-4o-mini
Temperature	Slider	No	Answer creativity (0-1)	0.7
Top K / Max Context Docs	Number	No	Documents in context (1-20)	5
Max Tokens	Number	No	Answer length limit (50-2000)	500
Similarity Threshold	Number	No	Minimum relevance	0.3
System Prompt	Text	No	Custom instructions	—
Include Sources	Toggle	No	Include source citations	—
Use Hybrid Search	Toggle	No	Enable BM25 + Vector	—
Answer Style	Select	No	comprehensive or concise	—

How It Works

1. Retrieve: Search the KB for relevant documents
2. Augment: Add documents to the AI context
3. Generate: AI creates an answer from the context

Question Example

Question: What is the return policy for electronics?

System Prompt (optional):
You are a customer support agent. Answer based only on the
provided documents. If the answer isn't in the documents,
say "I don't have that information."

Output Variables

{{kbragquery_1.answer}}            - AI-generated answer
{{kbragquery_1.sources}}           - Source documents used
{{kbragquery_1.confidence_score}}  - Answer confidence (0-1)
{{kbragquery_1.citations}}         - Specific citations
{{kbragquery_1.search_results}}    - Raw search results
{{kbragquery_1.success}}           - true/false

---

Text Response Node

Format and send AI-generated responses to users.

When to Use

• Send final responses to users
• Format AI output nicely
• Add typing indicators
• Control response timing

Configuration

Field	Description	Default
Text	Response template	{{output}}
Typing Indicator	Show typing before send	false
Delay	Seconds before sending	0

Formatting Options

Simple variable:

{{llmintegration_1.output}}

With formatting:

Here's what I found:

{{kbragquery_1.answer}}

Sources:
{{#each kbragquery_1.sources}}
- {{item.document_title}}
{{/each}}

Conditional content:

{{#if success}}
✅ Your request has been processed!
{{else}}
❌ Sorry, there was an error. Please try again.
{{/if}}

Array Iteration

Found {{results.length}} results:

{{#each results}}
{{@index1}}. {{item.title}}
   Relevance: {{item.similarity}}
{{/each}}

Output Variables

{{textresponse_1.response_text}} - Final formatted text
{{textresponse_1.output}}        - For chaining

---

Common AI Patterns

Pattern 1: Simple Q&A

User Message Trigger
    │
    ▼
KB RAG Query (question: {{message}})
    │
    ▼
Text Response ({{kbragquery_1.answer}})

Pattern 2: Contextual Response

User Message Trigger
    │
    ▼
Contact Lookup (get user context)
    │
    ▼
KB Search (find relevant docs)
    │
    ▼
Prompt Template (combine context + docs)
    │
    ▼
LLM Integration (generate response)
    │
    ▼
Text Response (send to user)

Pattern 3: Agent with Tools

User Message Trigger
    │
    ▼
Core Agent Node
    ├── Can use: KB Search
    ├── Can use: CRM Lookup
    ├── Can use: Email Send
    └── Can use: API Request
    │
    ▼
Text Response (agent's final answer)

Pattern 4: Content Generation

Database Trigger (new deal created)
    │
    ▼
Transform Data (prepare context)
    │
    ▼
LLM Integration (generate proposal)
    │
    ▼
Email Send (send to contact)

Pattern 5: Classification + Routing

Webhook Trigger (new inquiry)
    │
    ▼
LLM Integration (classify intent)
    │
    ▼
Multi-Condition (route by intent)
    │
    ├── Sales → Sales Workflow
    ├── Support → Support Workflow
    └── Billing → Billing Workflow

Compact Pattern Reference

Simple Generation:

[Prompt Template] → [LLM Integration] → [Text Response]

RAG-Enhanced Response:

[KB Search] → [LLM Integration with context] → [Text Response]

Agent Task:

[User Input] → [Core Agent] → [Response Builder]
                    ↓
              [Uses tools as needed]

---

Prompt Engineering Tips

Be Specific

❌ Bad: "Write an email"

✅ Good: "Write a professional follow-up email to a prospect
who attended our webinar yesterday. Include:
- Thank them for attending
- Summarize key points from the webinar
- Offer a free consultation
- Keep it under 150 words"

Use Examples

System Prompt:
Classify customer inquiries into categories.

Examples:
- "I need help resetting my password" → account_access
- "When will my order arrive?" → order_status
- "I want a refund" → billing

Now classify this inquiry:
{{customer_message}}

Set Constraints

Generate a product description.

Constraints:
- Exactly 2-3 sentences
- Include the key benefit
- End with a call-to-action
- No superlatives (best, greatest)
- Use active voice

Chain of Thought

Analyze this customer feedback and determine the sentiment.

Think step by step:
1. Identify the main topic
2. List positive mentions
3. List negative mentions
4. Weigh the overall tone
5. Provide final sentiment (positive/negative/neutral)

Feedback: {{feedback_text}}

tip

Reusable prompts can be stored and versioned in the Prompts Library, then referenced from the Prompt Template node. To measure and improve prompt quality, see AI Evaluation.

---

Cost Optimization

Choose the Right Model

Task Complexity	Recommended Model	Cost
Simple (classification, extraction)	gpt-4o-mini, claude-3-haiku	$
Medium (summarization, Q&A)	gpt-4o-mini, claude-3-sonnet	$$
Complex (reasoning, analysis)	gpt-4o, claude-3-opus	$$$

Reduce Token Usage

✅ Use concise prompts
✅ Limit context documents (max_context_docs)
✅ Set appropriate max_tokens
✅ Use temperature=0 for deterministic tasks

❌ Don't include unnecessary context
❌ Don't ask for verbose responses
❌ Don't use opus for simple tasks

Narrow Before You Generate

Use KB Search with metadata filters to narrow results before RAG queries — fewer, more relevant context documents mean lower cost and better answers.

---

Troubleshooting

Empty AI Response

1. Check prompt — Is it clear what you want?
2. Verify variables — Are they resolving?
3. Check max_tokens — Is it too low?
4. Review integration — Is the API key valid?

Poor Quality Answers

1. Improve prompt — Be more specific
2. Add examples — Show expected output
3. Use a better model — Upgrade if needed
4. Adjust temperature — Lower for factual tasks

KB Search Returns Nothing

1. Check KB content — Are documents indexed?
2. Lower threshold — Try a lower similarity score
3. Simplify query — Shorter, clearer queries
4. Enable hybrid search — Combine keyword + semantic

Agent Loops or Fails

1. Lower max_iterations — Prevent infinite loops
2. Simplify system prompt — Clearer instructions
3. Check tool availability — Are tools configured?
4. Review error logs — See what failed

---

Quick Reference

Variable Syntax

{{variable}}              - Simple variable
{{node.output.field}}     - Nested path
{{items[0].name}}         - Array access
{{var|"default"}}         - Fallback value
{{value|int}}             - Type conversion

Models (by capability)

High:    gpt-4o, claude-3-opus
Medium:  gpt-4o-mini, claude-3-sonnet
Fast:    gpt-3.5-turbo, claude-3-haiku

Temperature Guide

Value	Use Case
0.0	Deterministic (classification, extraction)
0.2-0.3	Balanced (Q&A, summaries)
0.7	Creative (content generation)
1.0+	Very creative (brainstorming)

KB Search Parameters

top_k: 20         - Candidates to consider
limit: 5          - Final results
threshold: 0.5    - Minimum similarity
bm25_weight: 0.5  - Hybrid balance

---

Next Steps

With AI nodes mastered, continue with:

1. Testing Workflows — Debug AI workflows
2. Error Handling — Handle AI failures gracefully
3. Knowledge Bases — Build the data your AI nodes search
4. Agentic Mode — Autonomous, skill-driven AI experiences