GitHub - sifaka-ai/sifaka: Sifaka is an open-source framework that adds reflection and reliability to large language model (LLM) applications.

# Clone the repository
git clone https://github.com/sifaka-ai/sifaka
cd sifaka

# Install with uv (recommended)
uv pip install -e .

# Or with standard pip
pip install -e .

# OpenAI (default provider)
export OPENAI_API_KEY=sk-...

# Or Anthropic
export ANTHROPIC_API_KEY=sk-ant-...

# Or Google
export GOOGLE_API_KEY=...

# Or Groq
export GROQ_API_KEY=...

from sifaka import improve
import asyncio

async def main():
    result = await improve("Write about renewable energy benefits")
    print(result.final_text)
    print(f"\nImprovement score: {result.improvement_score:.2f}")
    print(f"Iterations: {result.iteration}")

asyncio.run(main())

from sifaka import improve_sync

result = improve_sync("Write about renewable energy benefits")
print(result.final_text)

result = await improve("Your text")

# Access complete audit trail
for iteration in result.trace:
    print(f"Iteration {iteration.number}")
    print(f"  Critique: {iteration.critique}")
    print(f"  Improvement: {iteration.improvement}")
    print(f"  Time: {iteration.processing_time:.2f}s")

# OpenAI
result = await improve(text, provider="openai", model="gpt-4o-mini")

# Anthropic
result = await improve(text, provider="anthropic", model="claude-3-5-sonnet")

# Google
result = await improve(text, provider="google", model="gemini-1.5-flash")

# Groq (fast inference)
result = await improve(text, provider="groq", model="llama3-8b-8192")

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"])
    ]
)

from sifaka import improve

result = await improve("AI is important for business.")
print(result.final_text)
# Output: "Artificial intelligence transforms business operations by automating..."

from sifaka import improve
from sifaka.core.types import CriticType

# Single critic
result = await improve(
    "Explain quantum computing",
    critics=[CriticType.REFLEXION]
)

# Multiple critics
result = await improve(
    "Explain quantum computing",
    critics=[CriticType.REFLEXION, CriticType.SELF_REFINE]
)

from sifaka.critics.style import StyleCritic

result = await improve(
    "We offer comprehensive solutions for your needs.",
    critics=[StyleCritic(
        style_description="Casual and friendly",
        style_examples=["Hey there!", "No worries!"]
    )]
)

result = await improve(
    "The Great Wall of China is visible from space.",
    critics=[CriticType.SELF_RAG]
)
# Critiques factual accuracy and suggests corrections

result = await improve(
    "Guide on pest control methods",
    critics=[CriticType.CONSTITUTIONAL]
)
# Evaluates against safety principles

result = await improve(
    "Product launch announcement",
    critics=[CriticType.N_CRITICS]
)
# Gets feedback from technical expert, general audience, editor, skeptic perspectives

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

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

from sifaka import Config

config = Config(
    model="gpt-4",
    temperature=0.7,
    max_iterations=5,
    timeout_seconds=120
)

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

from sifaka.storage.file import FileStorage
from sifaka.storage.redis import RedisStorage

# File storage
result = await improve(
    "Your text",
    storage=FileStorage("./results")
)

# Redis storage
result = await improve(
    "Your text",
    storage=RedisStorage("redis://localhost:6379")
)

from sifaka.core.exceptions import ValidationError, CriticError

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

import asyncio

texts = ["Text 1", "Text 2", "Text 3"]
tasks = [improve(text) for text in texts]
results = await asyncio.gather(*tasks)

from sifaka.validators import BaseValidator

class CustomValidator(BaseValidator):
    async def validate(self, text: str) -> ValidationResult:
        # Your custom validation logic
        passed = "important_keyword" in text.lower()
        return ValidationResult(
            validator="custom",
            passed=passed,
            message="Must contain 'important_keyword'"
        )

result = await improve(text, validators=[CustomValidator()])

# Technical accuracy + readability
result = await improve(
    "Technical documentation",
    critics=[CriticType.REFLEXION, CriticType.STYLE]
)

# Safety + factual accuracy
result = await improve(
    "Health advice article",
    critics=[CriticType.CONSTITUTIONAL, CriticType.SELF_RAG]
)

# Comprehensive review
result = await improve(
    "Important business document",
    critics=[
        CriticType.SELF_REFINE,
        CriticType.N_CRITICS,
        CriticType.META_REWARDING
    ]
)

# LLM Provider Keys
OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=sk-ant-...
GOOGLE_API_KEY=...
GROQ_API_KEY=...

# Optional: Default settings
SIFAKA_DEFAULT_MODEL=gpt-4o-mini
SIFAKA_MAX_ITERATIONS=3
SIFAKA_TEMPERATURE=0.7

from sifaka import Config

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
)

┌─────────────────────────────────────────────┐
│           Sifaka Improvement Loop           │
└─────────────────────────────────────────────┘
                      │
                      ▼
        ┌──────────────────────────┐
        │   1. Generate/Modify     │
        │      (LLM Provider)      │
        └──────────────────────────┘
                      │
                      ▼
        ┌──────────────────────────┐
        │   2. Critique            │
        │   (Critics: Reflexion,   │
        │    Constitutional, etc)  │
        └──────────────────────────┘
                      │
                      ▼
        ┌──────────────────────────┐
        │   3. Validate            │
        │   (Validators: Length,   │
        │    Content, Custom)      │
        └──────────────────────────┘
                      │
                      ▼
        ┌──────────────────────────┐
        │   4. Improve             │
        │   (Apply Suggestions)    │
        └──────────────────────────┘
                      │
                      ▼
           [Repeat or Return Result]

# Run tests
pytest tests/

# Type checking
mypy sifaka/

# Linting
ruff check .

# Formatting
black .

# Coverage
pytest --cov=sifaka

@article{madaan2023self,
  title={Self-Refine: Iterative Refinement with Self-Feedback},
  author={Madaan, Aman and others},
  journal={arXiv preprint arXiv:2303.17651},
  year={2023}
}

@article{shinn2023reflexion,
  title={Reflexion: Language Agents with Verbal Reinforcement Learning},
  author={Shinn, Noah and others},
  journal={arXiv preprint arXiv:2303.11366},
  year={2023}
}

@article{bai2022constitutional,
  title={Constitutional AI: Harmlessness from AI Feedback},
  author={Bai, Yuntao and others},
  journal={arXiv preprint arXiv:2212.08073},
  year={2022}
}