AI Assistant in Pipeline Nodes

The AI Assistant feature enhances pipeline node configuration by providing intelligent, AI-powered content generation for text fields. It helps you create, refine, and improve prompts, code, templates, conditions, and descriptions directly within the pipeline editor.

---

What is AI Assistant?

AI Assistant is an integrated AI-powered tool that assists with content creation in pipeline nodes. Instead of manually writing complex prompts, code, or templates from scratch, you can:

• Generate content from natural language instructions
• Improve existing content by providing refinement prompts
• Iterate quickly with split-view comparison between original and improved versions
• Leverage AI reasoning to create better configurations

LLM Configuration

AI Assistant uses the pipeline's configured LLM model and settings. Configure this at the pipeline level to control which model, Creativity, Max Completion Tokens and other parameters are used for content generation.

---

Key Benefits

✔️ Faster Configuration - Generate complex content quickly from simple instructions

✔️ Higher Quality - Leverage AI to improve prompts, code, and templates

✔️ Iterative Refinement - Compare original and improved versions side-by-side

✔️ Context-Aware - AI Assistant understands the node type and purpose

✔️ Consistent Style - Uses your pipeline's LLM configuration

---

How AI Assistant Works

Access AI Assistant

AI Assistant is available on text input fields that support it. Look for the AI Assistant icon (🪄) on the right side of the field:

Click the icon to open the AI Assistant modal dialog.

AI Assistant Modal

The modal provides a full-screen editor with AI-powered content generation:

Modal Components:

1. Editor Area - View and edit your content with syntax highlighting
2. AI Prompt Input - Natural language instructions for content generation
3. Language Selector - Choose syntax highlighting (auto-detected)
4. Action Buttons - Copy, apply, and close controls

Single View Mode

When you first open AI Assistant with empty content or request new content:

• Editor displays your current content (or empty state)
• Prompt Input accepts your instructions
• Generate creates content based on your prompt
• Result appears directly in the editor

Split View Mode

When improving existing content:

• Left Panel - "Current Version" (read-only reference)
• Right Panel - "Improved Version" (AI-generated result)
• Compare - See differences side-by-side
• Apply - Replace current content with improved version
• Close - Discard improvements and return to single view

---

Supported Nodes and Fields

AI Assistant is available in the following pipeline nodes and their specific fields:

Node Type	Supported Fields	Purpose	AI Assistant Usage
LLM Node	System	System prompt defining LLM role and behavior	Generate system prompts for specialized agents
	Task	User task or query to process	Generate task descriptions and instructions
Code Node	Code	Python code to execute in sandbox	Generate Python code, improve existing code, add error handling, optimize algorithms
State Modifier Node	Jinja Template	Template for state transformation	Generate Jinja2 templates, create formatted output, build dynamic content
Router Node	Condition	Jinja2 template for routing logic	Generate routing conditions, create complex branching logic, build pattern matching rules
Decision Node	Description	Context for LLM decision-making	Generate decision criteria, create routing descriptions, define decision logic
Printer Node	Printer	Output formatting template	Generate display templates, create formatted messages, build user-facing output

---

LLM Node

Use Cases:

• Generate comprehensive system prompts for specialized agents
• Create clear, structured task instructions
• Build example conversations for testing

Generate System Prompt Example

Prompt: "Create a system prompt for an AI assistant that helps developers write unit tests"

Generated System Prompt:

You are an expert software testing assistant specializing in unit test creation. Your role is to:

1. Analyze code snippets provided by developers
2. Identify testable units and edge cases
3. Generate comprehensive unit tests with clear assertions
4. Follow best practices for the testing framework in use
5. Explain your testing strategy and coverage decisions

Provide tests that are:
- Clear and maintainable
- Cover happy paths and edge cases
- Include meaningful test names and descriptions
- Follow the AAA pattern (Arrange, Act, Assert)

---

Code Node

Use Cases:

• Generate data transformation logic
• Create calculations and business rules
• Build API integration code
• Improve code quality and error handling

Generate Tiered Discount Code

Prompt: "Write Python code to calculate tiered discounts: 10% for orders $100-500, 15% for $500-1000, 20% for $1000+"

Generated Code:

def calculate_discount(order_value):
    """
    Calculate tiered discount based on order value.

    Args:
        order_value (float): Total order value

    Returns:
        dict: Discount percentage and final price
    """
    # Get order value from state
    order_value = float(alita_state.get('order_value', 0))

    # Determine discount tier
    if order_value >= 1000:
        discount_rate = 0.20
    elif order_value >= 500:
        discount_rate = 0.15
    elif order_value >= 100:
        discount_rate = 0.10
    else:
        discount_rate = 0.0

    # Calculate discount and final price
    discount_amount = order_value * discount_rate
    final_price = order_value - discount_amount

    return {
        "discount_rate": discount_rate,
        "discount_amount": round(discount_amount, 2),
        "final_price": round(final_price, 2)
    }

# Execute and return results
result = calculate_discount(order_value)
return result

---

State Modifier Node

Use Cases:

• Format output messages with state variables
• Build reports and summaries
• Build structured responses

Generate User Story Template

Prompt: "Create a Jinja template that formats a user story with title, description, acceptance criteria, and metadata"

Generated Template:

# User Story: {{ title }}

## Description
{{ description }}

## Acceptance Criteria
{% for criterion in acceptance_criteria %}
- {{ criterion }}
{% endfor %}

## Metadata
- **Priority**: {{ priority }}
- **Story Points**: {{ story_points }}
- **Sprint**: {{ sprint }}
- **Assignee**: {{ assignee }}

---
*Created on {{ created_date }} by {{ author }}*

---

Router Node

Use Cases:

• Create multi-path routing logic
• Build approval workflows
• Implement status-based routing

Generate Router Condition

Prompt: "Create a router condition that routes to 'Approved' if status contains 'approved', 'Rejected' if 'rejected', 'Pending' if 'review', otherwise 'Default'"

Generated Condition:

{% if 'approved' in status|lower %}
Approved
{% elif 'rejected' in status|lower %}
Rejected
{% elif 'review' in status|lower or 'pending' in status|lower %}
Pending
{% else %}
Default
{% endif %}

---

Decision Node

Use Cases:

• Describe decision criteria clearly
• Provide context for intelligent routing
• Document decision-making logic

Generate Decision Description

Prompt: "Write a decision description for routing customer support tickets based on urgency and topic"

Generated Description:

Analyze the customer support ticket and route it to the appropriate team based on these criteria:

**High Priority (Critical Issues):**
- System outages or downtime
- Security vulnerabilities
- Payment processing failures
- Data loss or corruption

Route to: **CriticalSupport**

**Billing & Payments:**
- Invoice questions
- Payment issues
- Subscription changes
- Refund requests

Route to: **BillingTeam**

**Technical Support:**
- Feature questions
- Integration issues
- API problems
- Bug reports

Route to: **TechnicalSupport**

**General Inquiries:**
- Account questions
- Feature requests
- General questions

Route to: **GeneralSupport**

**Default:** Route to **GeneralSupport** if unsure

---

Printer Node

Use Cases:

• Display static completion messages
• Show formatted output combining text with variables
• Pause pipeline for user inspection
• Show progress updates during multi-step workflows

Generate Processing Summary (Type: F-String)

Create a comprehensive processing summary that combines multiple variables and includes conditional formatting.

Prompt: "Create a detailed processing summary message that shows the project name, total items processed, success count, failure count, success rate percentage, processing time, and a status indicator (✓ if success rate > 90%, ⚠ if 70-90%, ✗ if < 70%)"

Generated F-String:

==========================================
Processing Complete: {project_name}
==========================================

Results:
  • Total Items: {total_items}
  • Successful: {success_count} ({success_rate}%)
  • Failed: {failed_count}
  • Processing Time: {processing_time}s

Status: {status_indicator}

{completion_message}
==========================================

Example Output:

==========================================
Processing Complete: Customer Analytics Dashboard
==========================================

Results:
  • Total Items: 150
  • Successful: 142 (94.7%)
  • Failed: 8
  • Processing Time: 12.5s

Status: ✓

Processing completed successfully with high success rate.
==========================================

---

Using AI Assistant: Step-by-Step

1. Generate New Content

Scenario: You need to create a system prompt for an LLM node but aren't sure how to structure it.

Steps:

1. Open the LLM node configuration
2. Click the AI Assistant icon (🪄) on the "Task" field
3. Enter your instruction in the AI Prompt Input:

   Create a system prompt for a code reviewer that focuses on security best practices
   

4. Click "Generate" or press Enter
5. Review the generated content in the editor and edit if needed
6. Close the modal - content is now in the System field

2. Improve Existing Content

Scenario: You have a working prompt but want to make it more specific and detailed.

Steps:

1. Open AI Assistant on a field with existing content
2. Review current version in the editor
3. Enter improvement instruction in the AI Prompt Input:

   Make this prompt more specific for reviewing API security in REST endpoints
   

4. Click "Generate" - AI Assistant switches to split view
5. Compare versions side-by-side:
6. Left: Original content
7. Right: Improved content
8. Click "Apply" to use the improved version, or
9. Click "Close" (X) to discard and keep original

3. Iterative Refinement

Scenario: The generated content is good but needs additional refinement.

Steps:

1. Generate initial content using AI Assistant
2. Review the result in split view
3. Enter a refinement prompt without applying:

   Add examples of common security vulnerabilities to look for
   

4. Click "Generate" again - improved content updates
5. Continue refining with additional prompts:

   Make the tone more conversational and encouraging
   

6. Apply when satisfied with the final result

4. Error Handling and Retry

Scenario: AI Assistant encountered an error during generation.

What you'll see:

• Error message displayed in the editor
• Prompt input remains filled with your instruction
• Split view (if previously shown) stays open

Steps to retry:

1. Review the error message (e.g., "Error: Rate limit exceeded")
2. Your prompt is preserved - no need to retype
3. Click "Generate" again to retry
4. AI Assistant retries with the same prompt
5. Previous error is cleared if successful

---

Common Use Cases

Generate System Prompts for Specialized Agents

Challenge: Creating effective system prompts that properly constrain LLM behavior.

Solution with AI Assistant:

Prompt: "Create a system prompt for an AI assistant that generates SQL queries from natural language. It should only output valid SQL, never execute queries, and explain any assumptions."

Result: Detailed system prompt with role definition, constraints, output format, and safety guidelines.

Build Complex Routing Logic

Challenge: Creating multi-condition routing in Router nodes.

Solution with AI Assistant:

Prompt: "Create router condition for approvals: route to 'AutoApprove' if amount < 1000, 'ManagerApproval' if 1000-10000, 'ExecutiveApproval' if >10000, default to 'ReviewQueue'"

Result: Properly structured Jinja2 condition with numeric comparisons and fallback.

Create Data Transformation Code

Challenge: Writing Python code for state transformations.

Solution with AI Assistant:

Prompt: "Write Python code that takes a list of user dictionaries and creates a summary with total users, breakdown by role, and list of admin emails"

Result: Working Python code with error handling, state access, and structured output.

Format Output Templates

Challenge: Creating readable, well-formatted output displays.

Solution with AI Assistant:

Prompt: "Create a Jinja template for a daily report showing completed tasks, pending tasks, overdue tasks, and team member performance"

Result: Professional template with formatting, conditionals, and loops.

---

Best Practices

Writing Effective Prompts

✔️ Be Specific

• ✘ "Write a prompt"
• ✔️ "Write a system prompt for an AI assistant that reviews Python code for PEP 8 compliance"

✔️ Provide Context

• ✘ "Generate code"
• ✔️ "Generate Python code that reads CSV files, filters rows where status='active', and returns a list of user IDs"

✔️ Include Examples

• ✘ "Create a template"
• ✔️ "Create a Jinja template that displays user information in this format: 'Name: John Doe, Email: john@example.com, Role: Admin'"

✔️ Specify Format Requirements

• ✘ "Make routing logic"
• ✔️ "Create router condition using Jinja2 syntax that routes based on 'priority' field: high→Urgent, medium→Normal, low→Backlog"

Iterative Improvement

Start Simple, Refine

1. Generate basic content first
2. Review what AI created
3. Add refinement prompts incrementally
4. Apply when satisfied

Use Split View Effectively

• Keep original on left as reference
• Compare specific changes in improved version
• Apply only when improvements are clear
• Close split view to start over if needed

Content Validation

✓ Always Review Generated Content

• AI Assistant is powerful but not perfect
• Verify logic, syntax, and requirements
• Test generated code in safe environment
• Validate templates with sample data

✓ Understand Generated Code

• Don't blindly apply code you don't understand
• Review variable names and logic flow
• Ensure error handling is appropriate
• Check for security considerations

Performance Tips

Use Appropriate Field Types

• Use AI Assistant for complex, creative tasks
• For simple static values, type directly
• Reserve AI for prompts, code, templates

Leverage Pipeline LLM Configuration

• AI Assistant uses your pipeline's LLM settings
• Configure appropriate model for your needs
• Adjust temperature for creativity vs. consistency

---

Troubleshooting

AI Assistant Not Appearing

Issue: AI Assistant icon not visible on a field.

Possible Causes:

• Field doesn't support AI Assistant
• Pipeline LLM configuration is missing
• Node type doesn't enable AI Assistant

Solution:

1. Verify the field supports AI Assistant (see Supported Nodes)
2. Check pipeline configuration → Ensure LLM model is configured
3. Verify you're using a supported node type

Generated Content Not Applied

Issue: Clicking "Save" doesn't update the node field.

Possible Causes:

• Modal closed without saving
• Content validation failed
• Browser cached old value

Solution:

1. Re-open AI Assistant modal
2. Verify content is still in editor
3. Click "Save" button explicitly
4. Check node field updates after closing modal

Generation Fails with Error

Issue: AI Assistant displays error message instead of content.

Common Errors:

• "Rate limit exceeded" - Too many requests to LLM
• "Invalid API key" - LLM credentials issue
• "Timeout" - Request took too long

Solution:

1. Wait a moment and click "Generate" again (prompt preserved)
2. Simplify your prompt if too complex
3. Check pipeline LLM configuration and credentials
4. Contact administrator if persistent API errors

Split View Not Showing

Issue: Expected split view comparison but seeing single view.

Possible Causes:

• Current content is empty
• Current content starts with "Error"
• First-time generation

Expected Behavior:

• Empty content + Generate → Single view (new content)
• Existing content + Generate → Split view (comparison)
• Error content + Generate → Single view (retry)

---

Related Documentation

• LLM Node - Configure LLM nodes with AI Assistant
• Code Node - Write Python code with AI help
• State Modifier Node - Create Jinja templates
• Router Node - Build routing conditions
• Decision Node - Write decision descriptions
• Pipeline Configuration - Configure pipeline LLM settings
• States - Understanding pipeline state for AI-generated code

---