Download
Guardrails is an open source Python framework designed to help developers build more reliable and controlled applications powered by large language models. It provides mechanisms for validating and constraining both the inputs sent to a model and the outputs generated by it, helping reduce risks such as harmful content, prompt injection, or inaccurate responses. Guardrails works by applying configurable guards that intercept and evaluate interactions with the model before results are returned to the end user. These guards can detect and mitigate specific issues by applying validators that analyze content, enforce rules, or ensure structured output formats. Guardrails also supports generating structured data from language models, allowing developers to enforce schemas or type constraints on responses. A companion ecosystem known as a hub provides reusable validators that can be combined into input and output guards to address different reliability and safety concerns.
Features
- Input and output guards that intercept and validate LLM interactions
- Validator system for detecting risks such as unsafe or invalid content
- Structured data generation with schema or type validation
- Composable validation rules that can be combined for different policies
- Extensible ecosystem of reusable validators from a shared hub
- Automated corrective actions such as re-asking a model when validation fails
Project Samples
