guardrails-ai/guardrails

Features

• Python SDKs - Provides a Python library that wraps LLM calls with validation pipelines, corrective actions, and structured output generation.
• AI Observability Tracing - Logs model inputs, outputs, validator results, and performance metrics to debugging and observability platforms.
• LLM Provider Adapters - Wraps 100+ LLM providers behind a single API, enabling consistent guardrail application across models.
• LLM Provider Integrations - Connects to over 100 language models through a unified API to apply guardrails across providers.
• LLM Response Streaming - Streams validated LLM output incrementally, applying guardrails to each chunk as it arrives.
• LLM Risk Detectors - Detects and mitigates risks in model inputs and outputs with pre-built and custom validators.
• Error-Correction Feedback Loops - Automatically retries prompts or fixes outputs when validation fails, iterating until constraints are satisfied.
• Output Guardrails - Validates responses from any supported model against defined guardrails, returning structured and safe outputs.
• Structured Data Extraction - Generates structured, schema-compliant data from free-form model outputs using validation and re-prompting.
• LLM Output Constraints - Validates and fixes language model responses in real-time against custom rules and constraints.
• Schema-Enforced Output Parsers - Generates validated JSON or structured data from free-form model responses using user-defined schemas and function calling.
• LLM Schema Outputs - Returns validated model output as JSON or synthetic structured data, using schema models for reusable, team-wide definitions.
• LLM Structured Data Generators - Generates validated JSON or schema-based structured data from free-form model responses using function calling.
• LLM-to-Structured Data Converters - Formats free-form model responses into a user-defined schema using function calling or prompt optimization for reliable data extraction.
• Token Streaming - Receives model output token by token as it is generated for real-time display or processing.
• Custom Validation Rules - Ships a framework for building and registering new validators that detect specific risks or patterns beyond the built-in library.
• Unified Model Interfaces - Provides a single API pattern to call any of 100+ language models with consistent validation and error handling.
• Validation Pipelines - Validates LLM outputs through a sequential pipeline of reusable validators that can trigger corrective actions on failure.
• Model Interaction Monitors - Logs every model interaction and processing step automatically, exporting metrics to monitoring services.
• Agent Input and Output Validators - Validates and corrects responses from large language models against configurable safety and quality rules.
• LLM Risk Detectors - Intercepts inputs and outputs of large language models to identify, quantify, and mitigate specific types of risks using pre-built validators.
• AI Output Validation - Applies structural, type, and quality checks to model responses, enforcing rules like competitor or toxicity filters.
• LLM Stream Validators - Validates and corrects LLM output incrementally as tokens arrive for real-time feedback in interactive apps.
• Automatic Retries - Retries the LLM prompt or fixes the output automatically when a validation check fails.
• Guardrail Servers - Provides a Flask-based REST server that exposes guard validation logic through HTTP endpoints for remote clients.
• Custom Model Adapters - Provides a custom adapter function that integrates any unsupported language model into the guardrails workflow.
• Function Calling Interfaces - Passes structured tool definitions to the model and processes returned function calls within the guardrails pipeline.
• Guarded Tool Calls - Enables safe function calling with LLMs by applying guardrails to tool invocations and their responses.
• Guardrail Services - Exposes guard validation logic as a REST API or OpenAI-compatible endpoint for remote and cross-language use.
• LangChain Guardrail Integrations - Converts guardrail validators into LangChain-compatible runnable objects for use in chains and agents.
• Custom Adapters - Provides a custom adapter mechanism to integrate any unsupported LLM API into the guardrails workflow.
• LLM Observability - Traces and monitors model calls, validation results, and performance metrics in debugging and observability tools.
• Remote Guard Servers - Hosts guard validation logic on a server so clients in any language can enforce safety rules without running the framework locally.
• Cross-Model Validators - Applies the same validation logic to different language models, including major providers and integrations with popular frameworks.
• Guard Servers - Starts a Flask server that exposes guard functionality through HTTP endpoints for remote clients.
• LLM Interaction Concurrency - Scales LLM interactions by handling multiple calls concurrently with async support and parallelization.
• Streaming Corrections - Validates and corrects streaming LLM responses in real time for immediate feedback in interactive apps.
• Guard Servers - Exposes guard functionality as HTTP endpoints via Flask, allowing non-Python clients to enforce validation rules.
• Guard Configuration Files - Defines guard behavior in a config file that specifies validators and their violation responses, then loads it at server startup.
• Observability Instrumentation - Logs every model interaction, validator result, and performance metric for export to monitoring and debugging tools.
• Pipeline Integration - Converts guardrail validators into LangChain-compatible runnable objects for use in chains and agents.
• Reusable Validators - Defines reusable validators that check model outputs against custom rules, reducing code duplication and improving maintainability.
• Risk Rule Validators - Intercepts model outputs and checks them against pre-defined validators, raising exceptions or corrective actions when violations are detected.
• OpenAI-Compatible Servers - Ships an OpenAI-compatible server that routes LLM requests through input and output validation guards.
• Guardrails and AI Safety - Listed in the “Guardrails and AI Safety” section of the The Incredible Pytorch awesome list.
• Privacy and Safety - Tool for adding structure and quality guarantees to LLM outputs.
• Safety and Security - Adds safety guardrails to model interactions and outputs.
• Security & Privacy - Python framework for detecting risks and enforcing structured output.

Star history