NVIDIA gives chatbot developers programmable conversation rules

NVIDIA released NeMo Guardrails on 25 April, an open-source Python toolkit for defining how a chatbot handles conversations. Developers write rules in Colang, a language for dialogue flows, and can connect the toolkit to LangChain applications.

Photo: © 2020 Will Buckner, “Nvidia sign”, 31 October 2020. CC BY 2.0. Cropped; historical headquarters sign, not launch imagery. No endorsement implied.

The release’s architecture guide shows where those rules take effect. A model interprets the user’s intent; a predefined flow decides the next step when one matches. Otherwise, the system can ask the model what to do next, then generate a reply.

As NVIDIA’s technical explanation makes clear, the rules sit around the conversation. They complement model training, letting developers change an application’s behaviour through its dialogue flows.

The launch examples separate two checks: comparing an answer with retrieved reference material, and generating more answers to look for inconsistencies. Agreement among generated responses doesn’t establish that a statement is true.

There’s another boundary inside the configuration. The version 0.1.0 README warns that Colang files can call Python scripts and make multiple model calls, so files from untrusted sources need inspection. The rules themselves can trigger executable actions.

NVIDIA describes topical, safety and security goals for the toolkit. Its announcement doesn’t provide measured failure rates. The published mechanism gives developers programmable controls, rather than a blanket guarantee.