sifaka

Basic Usage

This guide covers the fundamental concepts and usage patterns of Sifaka.

Core Concepts

The Improvement Loop

Sifaka works through an iterative improvement loop:

1. Generate: Create or modify text
2. Critique: Analyze the text using various critics
3. Improve: Apply suggestions from critiques
4. Validate: Check if the text meets requirements
5. Repeat: Continue until satisfied or max iterations reached

Key Components

• Critics: Analyze text and provide feedback
• Validators: Check if text meets specific requirements
• Config: Controls behavior and parameters
• Result: Contains the final text and full audit trail

Simple Examples

For basic usage examples, see:

• Quickstart Guide - Getting started examples
• API Reference - Complete API documentation with examples

Using Different Critics

# Default critic (reflexion is used when none specified)
result = await improve("Explain quantum computing")

# Explicitly specify single critic
result = await improve(
    "Explain quantum computing",
    critics=["reflexion"]  # Same as default
)

# Multiple critics
result = await improve(
    "Explain quantum computing",
    critics=["reflexion", "constitutional", "self_rag"]
)

Adding Validators

from sifaka.validators import LengthValidator, ContentValidator

result = await improve(
    "Write a product description",
    validators=[
        LengthValidator(min_length=100, max_length=200),
        ContentValidator(required_terms=["features", "benefits"])
    ]
)

Controlling Iterations

# More iterations for higher quality
result = await improve(
    text="Draft email to client",
    max_iterations=5  # Default is 3
)

# Force improvements even if validation passes
result = await improve(
    text="Good text that passes validation",
    force_improvements=True
)

Understanding Results

The SifakaResult object contains:

# Access the final improved text
print(result.final_text)

# See all critiques
for critique in result.critiques:
    print(f"{critique.critic}: {critique.feedback}")

# Check validation results
for validation in result.validations:
    print(f"{validation.validator}: {'✓' if validation.passed else '✗'}")

# View generation history
for generation in result.generations:
    print(f"Generation: {generation.text[:100]}...")

# Get improvement summary
print(f"Iterations: {result.iteration}")
print(f"Improved: {result.final_text != result.original_text}")
print(f"Processing time: {result.processing_time:.2f}s")

Configuration

Using Config Object

from sifaka import Config

config = Config(
    model="gpt-4",
    temperature=0.7,
    max_iterations=5,
    force_improvements=True
)

result = await improve("Your text", config=config)

Common Configuration Options

config = Config(
    # Model settings
    model="gpt-4",              # LLM model to use
    temperature=0.7,            # Creativity (0.0-2.0)
    max_tokens=1000,            # Max response length

    # Critic settings
    critic_temperature=0.3,     # Lower = more consistent
    critic_context_window=3,    # Previous critiques to consider

    # Behavior settings
    max_iterations=3,           # Max improvement cycles
    force_improvements=False,   # Improve even if valid
    timeout_seconds=300,        # Overall timeout
)

Error Handling

from sifaka.core.exceptions import (
    ValidationError,
    CriticError,
    ModelProviderError
)

try:
    result = await improve(text)
except ValidationError as e:
    print(f"Validation failed: {e}")
except CriticError as e:
    print(f"Critic error: {e}")
except ModelProviderError as e:
    print(f"LLM error: {e}")
    print(f"Suggestion: {e.suggestion}")

Best Practices

1. Start Simple: Use default settings first
2. Choose Critics Wisely: Different critics for different tasks
3. Set Clear Validators: Define what “good” means
4. Monitor Costs: More iterations = more API calls
5. Handle Errors: Always wrap in try/except for deployment

Next Steps

• Learn about different critics
• Explore advanced usage
• Understand configuration options
• Create custom validators

This site is open source. Improve this page.