Player not loading? Watch on YouTube
This beginner lesson explains guardrails and observability for LLM applications through HR policy and Kubernetes chatbot examples. The speaker briefly introduces gateways for fallback routing and evaluations for checking model behavior.
The NeMo Guardrails demos cover off-topic requests, jailbreak attempts, sensitive information and harmful requests within an allowed topic. Scripted greetings illustrate how dialogue rules can answer routine messages without calling a model. The lesson separates input checks, output filtering and custom rules, including urgency detection. A prompt that changes the chatbot's persona bypasses the demonstrated system. The speaker also discusses Guardrails AI, Amazon Bedrock Guardrails and Azure AI Foundry; his claim that cloud guardrails never fail is not established by the demos.
The observability section explains spans, traces and waterfall views using model calls and agent workflows. Examples inspect token usage, model identity and execution time. The speaker recommends LangSmith for the LangChain ecosystem and describes Logfire for broader application tracing, including Python and API calls. The examples use hosted APIs.
