NeMo Guardrails
App : A product you use directly
NVIDIA's programmable guardrails to control dialogues, topics, and tool calls.
The most complete solution for conversational guardrails in enterprise settings; demanding to configure.
Security reflects where data goes, requested permissions and incidents on record at the verification date. It is not a guarantee of zero risk: check the deployment mode and the connections you enable.
What it does, concretely
Conversational applications built on language models can produce off-topic, inaccurate, or potentially harmful responses. NeMo Guardrails addresses this by letting developers insert a programmable control layer between their application code and the language model.
In practice, the tool relies on a dedicated configuration language called Colang, with a Python-like syntax. Developers define rules in text files: which topics are allowed or forbidden, which dialogue paths to follow, and how to react to specific user inputs. These rules apply at five distinct checkpoints: user input, dialogue, retrieved chunks (RAG), tool calls, and model output.
NeMo Guardrails is aimed at technical teams building conversational assistants, LLM endpoints, or enterprise RAG pipelines who need verifiable behavioural guarantees. It is compatible with several language models, including GPT-3.5, GPT-4, LLaMA-2, Falcon, and Mosaic, as well as with LangChain.
It is not suited for use without Python development skills: configuration requires writing code and Colang files, managing a Python 3.10 to 3.13 environment, and understanding one's LLM application architecture. Non-technical users will find no graphical interface for setting up rules.
What it offers
- ·Input rails: filter, reject, or transform the user message before any further processing.
- ·Dialog rails: steer the conversational path through Colang-defined flows, with predefined responses or conditional model invocation.
- ·Retrieval rails (RAG): filter or modify retrieved chunks before they are passed to the model.
- ·Execution rails: control the inputs and outputs of custom actions (tools) called by the model.
- ·Output rails: fact-checking, hallucination detection, and content moderation of generated responses before they reach the user.
- ·Built-in guardrails library: jailbreak and prompt injection detection, sensitive data masking, NVIDIA safety models, and third-party integrations.
- ·HTTP server exposing an API compatible with the OpenAI Chat Completions format, deployable via CLI or Docker.
- ·Compatibility with multiple language models and optional integration with LangChain.
In practice
- Price
- Free
- Commercial use
- Commercial use: yes
- Available on
- cli, docker
- Licence
- Apache-2.0
Security and data
- Rating
- Security: nothing to flag
- Your data
- Data stays local
- Maintenance
- Active
- Last release
- 16 September 2026
Links and repository checked on 25 October 2026 · added on 1 April 2023
How to install NeMo Guardrails
pip install nemoguardrails