Nvidia Corp. today announced today the release of new Nvidia Inference Microservices aimed at helping enterprise organizations develop artificial intelligence agents to address issues of trust, security and safety.

AI agents are a blossoming technology that is beginning to revolutionize how people interact with computers, but they also come with several critical issues. Agentic AI is currently set to change how knowledge workers accomplish tasks and customers “talk” to brands, but the large language models under the hood can still go off the rails and produce unwanted responses or create security concerns when malicious users break their safeguards.

Nvidia NIM is a set of containerized microservices designed to speed up the deployment of generative AI models and today’s announcement builds on NeMo Guardrails, a protective framework for developers that moderates AI models to allow them to build more secure, trustworthy AI agents.

Nvidia announced three NIM microservices that cover topic control, content safety and jailbreak protection. These microservices are highly optimized small, lightweight AI models that moderate responses from a larger model to improve application performance.

“One of the new microservices, built for moderating content safety, was trained using the Aegis Content Safety Dataset — one of the highest-quality, human-annotated data sources in its category,” said Kari Briski, vice president of enterprise AI models, software and services at Nvidia.

The data is curated and owned by Nvidia with more than 35,000 human-annotated data samples flagged for AI safety and jailbreak attempts to bypass system restrictions. The dataset will be publicly available on Hugging Face later this year.

The topic control NIM, for example, helps prevent agents from getting too “chatty” or diverging from their original mission by keeping them on topic. The longer a conversation goes with an AI chatbot, often the more likely it will begin to forget the original intent of the chat and the conversation can begin to wander, similar to how human conversations can tend to meander. Although this can be fine for people, it’s bad for chatbots, especially branded AI agents that might start talking about famous rock stars or competing products.

“Small language models, like those in the NeMo Guardrails collection, offer lower latency and are designed to run efficiently, even in resource-constrained or distributed environments,” said Briski. “This makes them ideal for scaling AI applications in industries such as healthcare, automotive and manufacturing, in locations like hospitals or warehouses.”

The NIM approach allows developers to stack multiple guardrails with minimal additional latency, or response time added. This is very important for most generative AI applications because customers don’t like waiting while watching three dots blinking or a spinning circle before text appears or a voice begins speaking.

“We’re always looking for ways to help associates go above and beyond for our customers,” said Chandhu Nair, senior vice president of data, AI and innovation at Lowe’s Cos. Inc. The home improvement retailer teamed up with Nvidia to assist customers and associates with AI. “With our recent deployments of Nvidia NeMo Guardrails, we ensure AI-generated responses are safe, secure and reliable, enforcing conversational boundaries to deliver only relevant and appropriate content.”

Image: Nvidia