Guardrails is a Python framework that adds input and output validation to large language model applications by detecting, quantifying, and mitigating specific types of risks.
The framework addresses the challenge of ensuring LLM outputs are safe and reliable. It works by running Input and Output Guards that intercept data flowing to and from language models. These Guards are composed of validators—pre-built risk detectors available through Guardrails Hub—that can be combined to address multiple risk categories simultaneously. The framework also helps generate structured data from LLMs by leveraging function calling where available or optimizing prompts to guide model output toward a desired schema defined through Pydantic models.
Guardrails suits teams building production LLM applications where output safety and consistency matter. It works well for applications requiring both risk mitigation and structured data generation. The tool can be deployed as a standalone Flask-based service with a REST API, making it suitable for microservice architectures. Developers can install validators directly from PyPI as standard packages rather than through a centralized package manager, giving flexibility in dependency management.
The project maintains an active benchmark comparing validator performance and latency across common risk categories. Development includes ongoing migration of the validator ecosystem toward standard package distribution, with clear communication about deprecation timelines for legacy infrastructure. The team publishes documentation and maintains a hub of pre-built validators that developers can discover and integrate into their applications.
