Step Library Overview

The Step Library is the heart of Jetty workflows, providing a comprehensive collection of reusable components that handle everything from AI model interactions to data processing, control flow, and evaluation. Each step is a self-contained function that can be configured, chained together, and orchestrated to create sophisticated AI/ML workflows.

What are Steps?

Steps are atomic units of work in Jetty workflows. They encapsulate specific functionality like:

• Calling AI models (OpenAI, Anthropic, Google Gemini, Replicate)
• Processing data (text manipulation, image processing, file operations)
• Controlling workflow execution (parallel processing, conditionals, loops)
• Evaluating outputs (LLM-as-judge, trajectory analysis, benchmarks)

Each step has:

• Input parameters - Configuration and data inputs
• Output values - Results that can be passed to other steps
• Dependencies - Clear ordering and data flow requirements
• Error handling - Robust failure management and recovery

Step Categories

The Step Library is organized into functional categories:

🤖 AI Models (15 steps)

Modern AI model integrations with unified configuration patterns:

• Google Gemini (5 steps) - Text generation, JSON/text reading, file processing, image generation
• LiteLLM (6 steps) - Multi-provider access to 100+ models (OpenAI, Claude, etc.)
• Replicate (8 steps) - Image generation, video generation, text streaming, embeddings

🔀 Control Flow (5 steps)

Orchestration and parallel processing:

• List Emit Await - Fan-out parallel workflow processing
• Extract From Trajectories - Aggregate results from child workflows
• Number Sequence Generator - Generate numeric sequences
• Conditional Branch - Conditional logic
• Loop Counter - Iteration state management

🔧 Data Processing (7 steps)

File manipulation, downloads, and integrations:

• Tools - Text operations, file I/O, image metadata, webhooks

⚖️ Evaluation (3 steps)

Assessment and scoring frameworks:

• Simple Judge - LLM-as-judge evaluation
• Select Trajectories - Trajectory filtering and selection
• Visualize Correlation - Statistical analysis and plotting

🧪 Development (3 steps)

Testing and development utilities:

• Text Echo - Echo text input (for testing)
• Text Doubler - Double text output (for testing)
• Random Compliance Check - Generate random scores (for testing)

Step Discovery by Use Case

Text Generation

• gemini_prompt - Google Gemini text generation
• litellm_chat - Multi-provider chat completions
• replicate_text_stream - Streaming text generation

Image Generation

• gemini_image_generator - Gemini image generation
• litellm_image_generation - DALL-E and compatible models
• replicate_text2image - Flux, Stable Diffusion, and more

Video Generation

• replicate_text2video - Seedance video generation

Image Analysis

• litellm_vision - Vision model analysis
• replicate_extract_embeddings_url - CLIP embeddings

Document Processing

• gemini_file_reader - Multi-format file analysis
• gemini_json_reader - JSON data analysis
• read_text_file - Text file reading
• text_concatenate - Combine multiple texts

Evaluation & Scoring

• simple_judge - LLM-as-judge evaluation
• select_trajectories - Filter and select trajectories
• visualize_correlation - Statistical analysis

Parallel Processing

• list_emit_await - Fan-out to child workflows
• extract_from_trajectories - Collect results

Benchmarking

• harbor_terminal_bench - Agent terminal benchmarks
• swe_bench_docker_eval - SWE-bench evaluation

Configuration Patterns

All steps follow consistent configuration patterns:

Common Parameters

{
  "activity": "step_name",
  "model": "model_identifier",
  "temperature": 0.7,
  "max_tokens": 1000
}

Secrets Management

Steps integrate with Jetty's secrets management:

• Environment variables (development)
• Organization-scoped secrets (production)
• Provider-specific API keys

Data Flow

Steps connect through output → input references:

{
  "prompt": "previous_step.outputs.generated_text",
  "input_data": "init_params.user_data"
}

Environment Variables

Steps automatically use these environment variables for API authentication:

Variable	Used By
OPENAI_API_KEY	LiteLLM steps
ANTHROPIC_API_KEY	LiteLLM steps
GEMINI_API_KEY	Gemini steps
REPLICATE_API_TOKEN	Replicate steps
LITELLM_API_KEY	LiteLLM proxy

Quick Links

Getting Started

• Quick Start Guide - Set up your first workflow
• First Flow Tutorial - Build a complete workflow

By Provider

• Google Gemini - Native Google AI integration
• LiteLLM - 100+ model providers
• Replicate - Specialized and community models

By Function

• Control Flow - Orchestration patterns
• Evaluation - LLM-as-judge

Best Practices

1. Start Simple - Begin with single-step workflows and gradually add complexity
2. Use Path Expressions - Reference outputs from previous steps dynamically
3. Leverage Secrets - Store API keys securely using secrets management
4. Handle Errors - Check step outputs for success/error indicators
5. Monitor Costs - Use appropriate models for task complexity
6. Test Incrementally - Validate individual steps before chaining

Next Steps

• Explore AI Models for text and image generation
• Learn about Control Flow for parallel processing
• Review Evaluation for LLM-as-judge patterns
• Check Guides for practical tutorials