NVIDIA-NeMo/Guardrails

Features

• AI Safety Guardrails - Provides an open-source framework for defining and enforcing safety policies on LLM inputs, outputs, and dialog flows.
• Guardrail Servers - Ships a dedicated HTTP server that exposes guardrail configurations via a chat completion endpoint.
• Guardrailed Inference Executions - Provides a Python API and server to run LLM inference with enforced safety guardrails.
• Guardrail-Enforced - Exposes an HTTP server with a chat completion endpoint that enforces guardrail policies.
• Conversation Flow Design - Defines and enforces conversational paths and interaction patterns for multi-turn LLM conversations.
• Dialog Flow Engines - Ships a Colang-based runtime engine for steering multi-turn LLM conversations through predefined interaction patterns.
• Conversational Dialog Flow Languages - Defines conversational guardrails using a custom modeling language for dialog states and transitions.
• Input Filters - Applies checks to user messages before they reach the LLM, allowing rejection or alteration of harmful input.
• LLM Provider Integrations - Integrates with OpenAI, Anthropic, HuggingFace, and NVIDIA NIM through a unified interface.
• Provider Abstraction Layers - Wraps multiple model providers behind a common interface that normalizes API differences.
• LLM Input Safety Interceptions - Intercepts user inputs before they reach the model and blocks those violating defined safety or security policies.
• Agent Steering Languages - Uses a modeling language to define conversational paths and enforce predefined interaction patterns.
• Conversational Dialog Guardrails - Enforces conversational policies that control the flow and content of multi-turn LLM interactions.
• Middleware-Style Guardrail Pipelines - Processes LLM requests and responses through a sequential chain of guardrail modules.
• Content Moderation - Inspects generated responses and can reject or modify them to remove unsafe, sensitive, or off-topic content.
• Security Policy Enforcers - Enforces configurable safety and security policies on LLM inputs and outputs to block harmful content.
• Model Safety Filters - Inspects model responses against configurable rules and blocks or modifies content that violates safety policies.
• Guardrail Policy Configurations - Declares guardrail rules, model connections, and dialog flows in YAML files parsed at runtime.
• LLM Behavior Policy Configurations - Defines LLM behavior policies using YAML configuration files and Colang dialog flows.
• Guardrail Plugin Architectures - Organizes safety checks as composable plugins that intercept inputs and outputs in the LLM pipeline.
• Guardrail Execution Traces - Traces which rails activate during conversations and monitors LLM calls to debug flow execution and identify bottlenecks.
• Guarded Tool Calls - Inspects inputs and outputs of custom actions called by the LLM to enforce safety and security constraints.
• Guardrail Reliability Evaluation - Measures the accuracy and performance of safety, fact-checking, moderation, and hallucination guardrails.
• LangChain Guardrail Integrations - Connects safety guardrails to LangChain chains, runnables, and LangGraph workflows.
• LLM Tracing Systems - Captures user inputs and model outputs in trace events for detailed analysis with a configurable privacy toggle.
• GenAI Trace Exporters - Exports guardrail execution traces using OpenTelemetry semantic conventions for GenAI observability.
• RAG Context Pruning - Inspects and filters retrieved chunks before they are used to prompt an LLM in RAG pipelines.
• Context Validators - Checks and filters retrieved chunks before they are used to prompt an LLM, preventing unsafe context injection.
• RAG Context Validations - Intercepts retrieved document chunks and applies configurable checks before they reach the LLM prompt.
• Production Deployments - Runs guardrail configurations as an API server, inside Docker, or as microservices for remote use.
• OpenTelemetry-Integrated Monitors - Emits traces, logs, and metrics via OpenTelemetry for end-to-end observability into guardrail execution.
• OpenTelemetry Exporters - Exports guardrail execution traces using OpenTelemetry semantic conventions for GenAI observability.
• Security and Sandboxing - Toolkit for adding programmable safety guardrails.

سجل النجوم