Sifaka

Simple AI text improvement through research-backed critique with complete observability

What is Sifaka?

Sifaka improves AI-generated text through iterative critique using research-backed techniques. Instead of hoping your AI output is good enough, Sifaka provides a transparent feedback loop where AI systems validate and improve their own outputs.

Core Value: See exactly how AI improves your text through research-backed techniques with complete audit trails.

Installation

# Install from PyPI
pip install sifaka

# Or with uv
uv pip install sifaka

Quick Start

1. Set up your API key

Sifaka requires an LLM API key. Set one of these environment variables:

export OPENAI_API_KEY="your-api-key"     # For OpenAI (GPT-4, etc.)
# or
export ANTHROPIC_API_KEY="your-api-key"  # For Claude
# or
export GEMINI_API_KEY="your-api-key"     # For Google Gemini
# or
export GROQ_API_KEY="your-api-key"       # For Groq
# or (for local Ollama - no API key needed)
export OLLAMA_BASE_URL="http://localhost:11434/v1"  # Optional, defaults to localhost

Or create a .env file in your project:

OPENAI_API_KEY=your-api-key

Using Ollama (Local LLMs):

from sifaka import improve_sync, Config
from sifaka.core.config import LLMConfig

# Use Ollama with specific model (must set critic_model too!)
config = Config(
    llm=LLMConfig(
        provider="ollama",
        model="mistral:latest",
        critic_model="mistral:latest"  # Important: set this to use Ollama for critiques
    )
)
result = improve_sync("Climate change is bad.", config=config)

2. Use Sifaka

from sifaka import improve_sync

# Simple one-liner
result = improve_sync("Climate change is bad.")
print(result.final_text)

📚 Full Documentation →

Key Features

• 🔬 Research-Backed: Implements Reflexion, Constitutional AI, Self-Refine, and more
• 👁️ Complete Observability: Full audit trail of every improvement
• 🎯 Simple API: One function does everything you need
• 💾 Memory-Safe: Bounded history prevents memory leaks
• ⚡ Fast: Minimal dependencies, maximum performance

Documentation

Getting Started

• Installation - Installation options and setup
• Quickstart Guide - Get up and running in 5 minutes
• Basic Usage - Common usage patterns
• API Reference - Complete API documentation

User Guides

• Critics Guide - Available critics and usage
• Validators Guide - Input validation options
• Configuration - Configuration options
• Advanced Usage - Advanced patterns

Architecture & Development

• Architecture Overview - System design
• Contributing - How to contribute
• Changelog - Version history

Examples

• Working Examples - Code examples for all features

Research Foundation

Sifaka implements these peer-reviewed techniques:

• Reflexion - Self-reflection for iterative improvement
• Constitutional AI - Principle-based evaluation
• Self-Refine - Iterative self-improvement
• N-Critics - Ensemble of diverse perspectives
• Self-RAG - Retrieval-augmented critique
• Meta-Rewarding - Two-stage meta-evaluation
• Self-Consistency - Multiple reasoning paths
• Self-Taught Evaluator - Contrasting outputs with reasoning traces
• Agent4Debate - Multi-agent competitive debate dynamics
• Style - Transform text to match specific writing styles and voices
• Prompt - Simple prompt-engineered critic for simple use cases

Development

git clone https://github.com/sifaka-ai/sifaka
cd sifaka
pip install -e ".[dev]"
pytest

See CONTRIBUTING.md for development guidelines.

License

MIT License - see LICENSE file for details.