AI reliability: NVIDIA NEMO Guardrails integrates trustlab’s reliable language models.

As a company adopts LLMS (Lange Language Models) in the application, there is an urgent problem: the creation of misunderstandings or wrong output, often called ‘hallucinations’. To solve this problem, NVIDIA aims to provide a powerful solution to improve AI reliability by integrating Cleanlab’s reliable language model (TLM) into the Nemo Guardrails platform.

NVIDIA NEMO GUARDRAILS Overview

NVIDIA NEMO Guardrails is a comprehensive platform designed to implement AI policy in the AI ​​application. It provides extended frameworks for guaranteeing content safety, detecting potential jailbreak, and controlling conversation topics. The platform integrates both NVIDIA’s monopoly safety mechanisms and third -party solutions to provide a unified approach to AI safety.

For example, Nemo GuardRails uses LLM’s own checks with tools such as NVIDIA’s LLAMA 3.1 Nemoguard Nemoguard Safety Nim and LLAMA GUARD of Meta. This tool immediately displays a violation by conducting a real -time audit of AI production text for pre -defined policies. The platform also supports integration with external guard rails such as ActiveScore of ActiveFence to improve flexibility and comprehensive sexuality.

Clean the overview of a reliable language model

If you integrate the reliable language model of Cleanlab into Nemo Guardrails, there is a significant development of AI safety. TLM records the reliability of LLM output through advanced uncertainty estimation technology. This feature is important for applications such as customer support systems that can expand AI production responses to human agents if they are considered unreliable.

TLM is advantageous in a scenario that requires RAG (RAG), especially in the flag. It supports automated LLM systems when classifying information and running tool calls with greater stability.

Actual Application: Customer Support AI Assistant

NVIDIA has developed AI Assistant, a customer support of the e -commerce platform, to show the integration with TLM with Nemo Guardrails. The assistant uses the company’s policy as a guide for the situation to handle inquiries about delivery, returns and refunds.

In fact, when a customer querys a return policy for the product, the AI ​​assistant refer to the policy to match the documented guidelines. If the response seems incomprehensible, the TLM prompts the system to provide a polar back response or escalate the query to a human agent.

Evaluation and implementation

In a variety of customer support scenarios, GuardRails showed the ability to effectively detect and manage hallucinations. For example, when asked about refunds for inaccurate items, the AI ​​assistant provided a high -reliable response while closely complying with policy guidelines.

On the contrary, if a policy, such as inquiries about returning certain types of jewelry, is ambiguous, Guardrails has been potentially misleading and decided to expand the human review problem.

The implementation of these GuardRails assesses the reliability of AI responses by constructing the Nemo Guardrails framework to utilize the TLM API of Cleanlab. Based on reliability scores, the system decides whether or not to respond to the user.

conclusion

Integrating the reliable language model of NVIDIA’s Cleanlab into Nemo Guardrails provides a powerful solution to improve the reliability of AI applications. By solving the challenge of hallucinations, this collaboration provides a tool for developers to build a safer and more reliable AI system. The participation of Cleanlab in NVIDIA’s Inception Program emphasizes promise to AI technology and innovation development.

Image Source: Shutter Stock