strux 0.1.1

Strux is a Python framework for structured outputs model versioning

Strux

Strux is a Python library for building type-safe regression testing pipelines for structured outputs. It enables developers to easily validate model outputs against expected schemas and thresholds with minimal boilerplate.

Key Features

• 🔒 Type-Safe: Built on Pydantic for robust schema validation and type checking
• 🔄 Flexible Pipelines: Chain multiple inference steps with automatic schema validation
• 📊 Configurable Validation: Define strict, relaxed, or custom validation strategies
• 🔌 Extensible: Built-in PostgreSQL support with easy extension for other data sources
• 📈 Rich Results: Detailed validation reports with field-level insights

Quick Start

from strux import Sequential, RegressionConfig, ValidationLevel
from pydantic import BaseModel

Define your Schemas

class InputSchema(BaseModel):
    value: float
    text: str

class OutputSchema(BaseModel):
    processed: bool
    confidence: float

Define your Inference Function

def process_text(data: InputSchema) -> OutputSchema:
    return OutputSchema(
    processed=data.text.upper(),
    confidence=data.value 100
)

Create and run pipeline

pipeline = Sequential.from_steps(
    data_source=your_data_source,
    steps=[
        ("process", process_text, OutputSchema)
    ],
config=RegressionConfig(
    OutputSchema,
    strict_fields=["processed"],
    relaxed_fields=["confidence"]
    )
)
results = pipeline.run()

Installation

pip install strux

Why Strux?

• Type Safety: Catch schema mismatches early with Pydantic validation
• Flexible Validation: Configure different validation levels per field
• Pipeline Composition: Chain multiple inference steps with automatic validation
• Production Ready: Built-in support for batch processing and error handling

Documentation

For detailed documentation, see the docs/ directory:

• Getting Started
• Core Concepts
• API Reference
• Examples

Contributing

Contributions are welcome! Please read our Contributing Guidelines for details.