kaygraph 0.3.1

A context-graph framework for building production-ready AI applications.

KayGraph

A domain-specific language (DSL) for building context-aware AI applications.

---

📚 New to KayGraph?

Choose your path:

You are...	Start here
👤 Human Developer	Follow the 10-minute quickstart below
🤖 AI Coding Agent	Load LLM_CONTEXT_KAYGRAPH_DSL.md first
🎯 Task-focused	"I need to build X" → QUICK_FINDER.md
📖 Exploring all examples	Browse 70 workbooks in 16 categories
⚠️ Debugging/stuck	Check Common Patterns & Errors

---

What is KayGraph?

KayGraph is an opinionated DSL for expressing business problems as AI agent pipelines. Think of it as building blocks for AI workflows - a 500-line core that provides powerful abstractions without the bloat.

Core Philosophy:

• 🎯 DSL-First: Express complex AI workflows declaratively
• 🪶 Zero Dependencies: Pure Python standard library (500 lines of core code)
• 🔧 Bring Your Own Tools: Works with any LLM, database, or service
• 📦 Production-Ready: 70 battle-tested examples across 16 categories

---

Quick Start

Installation

# Using pip
pip install kaygraph

# Or from source
git clone https://github.com/KayOS-AI/KayGraph.git
cd KayGraph
pip install -e .

Your First KayGraph Workflow

from kaygraph import Node, Graph

class AnalyzeNode(Node):
    def prep(self, shared):
        """Phase 1: Read from shared context"""
        return shared.get("text")

    def exec(self, text):
        """Phase 2: Execute logic (LLM call, API, etc.)"""
        return f"Analyzed: {text}"

    def post(self, shared, prep_res, exec_res):
        """Phase 3: Write results back to shared context"""
        shared["result"] = exec_res
        return None  # End of workflow

# Build and run
analyze = AnalyzeNode()
graph = Graph(analyze)

shared = {"text": "Hello KayGraph!"}
graph.run(shared)
print(shared["result"])  # "Analyzed: Hello KayGraph!"

That's it! Three phases: prep() → exec() → post()

---

For Humans: Learning Path

1. Start with the Basics (10 minutes)

# Try the simplest example
cd workbooks/01-getting-started/kaygraph-hello-world
python main.py

2. Explore by Use Case

Use the task-based finder to jump to what you need:

📋 workbooks/QUICK_FINDER.md - "I need to build..."

• An AI Agent → Examples + patterns
• A Chatbot → Chat patterns
• A RAG System → Retrieval patterns
• Batch Processing → Data pipeline patterns
• Production API → Deployment examples

3. Browse All 70 Examples

📚 workbooks/WORKBOOK_INDEX_CONSOLIDATED.md - Complete catalog

16 Categories:

1. Getting Started (1)
2. Core Patterns (2)
3. Batch Processing (5)
4. AI Agents (9)
5. Workflows (12)
6. AI Reasoning (4)
7. Chat & Conversation (4)
8. Memory Systems (3)
9. RAG & Retrieval (1)
10. Code Development (2)
11. Data & SQL (4)
12. Tools Integration (7)
13. Production & Monitoring (8)
14. UI/UX (4)
15. Streaming & Realtime (2)
16. Advanced Patterns (2)

---

For Coding Agents: DSL Reference

🤖 LLM_CONTEXT_KAYGRAPH_DSL.md - Complete DSL specification for AI agents

This document contains everything a coding agent needs to:

• Understand the 3-phase node lifecycle
• Build graphs with proper action routing
• Use all node types (Async, Batch, Parallel, Validated, Metrics)
• Follow production patterns
• Avoid common anti-patterns

For AI Assistants (Claude, GPT-4, etc.):

Load the LLM_CONTEXT_KAYGRAPH_DSL.md file to understand KayGraph's
domain-specific language and generate production-ready code.

---

Core Concepts (5-Minute Overview)

The 3-Phase Node Lifecycle

Every node follows this pattern:

class MyNode(Node):
    def prep(self, shared):
        """
        Phase 1: READ from shared store
        - Gather data needed for execution
        - Access shared context
        - Return data for exec()
        """
        return shared.get("input_data")

    def exec(self, prep_res):
        """
        Phase 2: EXECUTE logic (NO shared access!)
        - Process data (LLM calls, APIs, etc.)
        - Pure function - can be retried
        - Return results
        """
        return process_data(prep_res)

    def post(self, shared, prep_res, exec_res):
        """
        Phase 3: WRITE to shared store and route
        - Update shared context with results
        - Return action string for routing
        - Return None for default/end
        """
        shared["output"] = exec_res
        return "next_action"  # or None

Why this matters:

• prep() and post() have context, exec() is pure
• exec() can be retried independently (resilience!)
• Clear separation of concerns

Graph Composition

# Chain nodes with default flow
node1 >> node2 >> node3

# Named actions for branching
decision_node >> ("approve", approval_node)
decision_node >> ("reject", rejection_node)

# Complex workflows
extract >> transform >> ("validate", validator)
validator >> ("success", loader)
validator >> ("failed", error_handler)

Shared Store Pattern

# Simple dictionary for context
shared = {
    "user_id": "123",
    "input": "Analyze this text",
    "history": []
}

# Nodes read and write to it
graph.run(shared)

# Results available after execution
print(shared["analysis_result"])

---

Key Features

Node Types

Type	Use Case	Example
Node	Standard sync operations	API calls, file I/O
AsyncNode	I/O-bound async operations	Concurrent API calls
BatchNode	Process iterables	Process 1000 records
ParallelBatchNode	Concurrent batch processing	Parallel data transforms
ValidatedNode	Input/output validation	Production pipelines
MetricsNode	Performance tracking	Monitoring, profiling

Production Features

• ✅ Retry Logic: Built-in with max_retries and wait
• ✅ Fallback Handling: exec_fallback() for graceful degradation
• ✅ Validation: Input/output type checking
• ✅ Metrics: Execution time, retry counts, success rates
• ✅ Logging: Comprehensive debug support
• ✅ Context Managers: Resource cleanup

---

Example Patterns

Agent Pattern

# Decision-making loop
think >> analyze >> ("use_tool", tool_node)
analyze >> ("respond", response_node)
tool_node >> think  # Loop back for reasoning

RAG Pattern

# Offline indexing
extract >> chunk >> embed >> store

# Online retrieval
query >> search >> rerank >> generate

Workflow Pattern

# Human-in-the-loop
process >> review >> ("approve", execute)
review >> ("reject", notify)
review >> ("modify", process)  # Loop back

---

Common Use Cases

I want to build...	Start here	Combine with
ChatGPT Clone	chat-memory	streaming-llm + chat-guardrail
Research Assistant	agent	rag + tool-search + agent-tools
Data Pipeline	workflow	batch + validated-pipeline
Multi-Agent System	multi-agent	supervisor + agent-memory
Production API	production-ready-api	metrics-dashboard + fault-tolerant

See workbooks/QUICK_FINDER.md for the complete list.

---

Scaffolding Tool

Generate production-ready boilerplate instantly:

# Generate a basic node
python scripts/kaygraph_scaffold.py node DataProcessor

# Generate an agent
python scripts/kaygraph_scaffold.py agent ResearchBot

# Generate a RAG system
python scripts/kaygraph_scaffold.py rag DocumentQA

# Generate a chat application
python scripts/kaygraph_scaffold.py chat CustomerSupport

# See all templates
python scripts/kaygraph_scaffold.py --help

Each template includes:

• Complete working code
• Documentation with TODOs
• requirements.txt with optional dependencies
• README with quickstart

---

Documentation

For Developers

• 📘 CLAUDE.md - Development guide for human developers and AI assistants
• 📕 COMMON_PATTERNS_AND_ERRORS.md - Avoid common mistakes and follow best practices
• 📗 docs/ - Architecture, patterns, and best practices
• 📙 CHANGELOG.md - Version history

For AI Coding Agents

• 🤖 LLM_CONTEXT_KAYGRAPH_DSL.md - Complete DSL reference
• 📝 workbooks/WORKBOOK_INDEX_CONSOLIDATED.md - All 70 examples organized
• ⚠️ COMMON_PATTERNS_AND_ERRORS.md - Common errors and how to avoid them

For Quick Tasks

• 🎯 workbooks/QUICK_FINDER.md - Task-based navigation ("I need to build...")
• 🚀 workbooks/guides/LLM_SETUP.md - Set up local LLMs with Ollama

---

Why KayGraph?

The 500-Line Philosophy

KayGraph's core is intentionally 500 lines. This isn't a limitation - it's a feature.

Why?

• ✅ You can read and understand the entire framework in one sitting
• ✅ No hidden magic - just Python classes and composition
• ✅ Easy to debug - it's just your code
• ✅ No vendor lock-in - bring your own LLM, database, tools
• ✅ Production-ready patterns without framework bloat

When humans can specify the graph, AI agents can automate it.

Zero Dependencies

The core framework has zero external dependencies. All examples that use LLMs, databases, or other services provide implementation templates - you bring your own tools.

This means:

• 🪶 Tiny install footprint
• 🔧 Total control over your stack
• 🎯 Only pay for what you use
• 🚀 No dependency hell

---

Project Structure

KayGraph/
├── kaygraph/                      # Core framework (500 lines!)
│   └── __init__.py                # All abstractions in one file
│
├── workbooks/                     # 70 production examples
│   ├── 01-getting-started/        # Start here
│   ├── 04-ai-agents/              # Agent patterns
│   ├── 09-rag-retrieval/          # RAG systems
│   └── ...                        # 13 more categories
│
├── scripts/                       # Scaffolding tools
│   └── kaygraph_scaffold.py       # Generate boilerplate
│
├── docs/                          # Comprehensive guides
├── tests/                         # Unit tests
│
├── LLM_CONTEXT_KAYGRAPH_DSL.md   # For coding agents
├── CLAUDE.md                      # For developers
└── README.md                      # You are here

---

Testing & Quality

All 70 workbooks are validated for:

• ✅ Valid structure (README.md + main.py)
• ✅ Valid Python syntax
• ✅ All imports resolve
• ✅ 100% pass rate

Run validation yourself:

python tasks/workbook-testing/validate_all_workbooks.py

---

Contributing

We welcome contributions! Please see our Contributing Guide for details.

Quick Contribution Ideas

• 🐛 Fix bugs in examples
• 📝 Improve documentation
• 💡 Add new workbook examples
• 🧪 Expand test coverage
• 🎨 Enhance scaffolding templates

---

Community & Support

• 🐛 Bug Reports: GitHub Issues
• 💬 Discussions: GitHub Discussions
• 📧 Email: [Your contact email]

---

License

MIT License - see LICENSE for details.

---

Quick Reference Card

# Node Lifecycle
class MyNode(Node):
    def prep(self, shared):      # 1. Read context
        return data
    def exec(self, prep_res):    # 2. Execute (pure!)
        return result
    def post(self, shared, prep_res, exec_res):  # 3. Write & route
        shared["result"] = exec_res
        return "action"  # or None

# Graph Building
node1 >> node2                   # Default flow
node1 >> ("action", node2)       # Named action
node1 >> node2 >> node3          # Chain

# Running
graph = Graph(start_node)
shared = {"input": "data"}
graph.run(shared)
print(shared["output"])

---

Built with ❤️ by the KayOS Team

Ready to build? Start with workbooks/01-getting-started/kaygraph-hello-world