nemo-guardrails

Install nemoguardrails, define safety flows in Colang with RailsConfig.from_content, then wrap your LLM with LLMRails and call rails.generate on user messages. The rails intercept unsafe input before it reaches the model.

Define a Colang flow matching jailbreak phrases like "Ignore previous instructions" and route them to a refusal response. You can also add LLM-based scoring and tune the threshold to reduce false positives.

NeMo Guardrails is a programmable framework orchestrating multiple safety checks at runtime, while LlamaGuard is a standalone moderation model. The two combine well: register LlamaGuard as an input and output check inside a NeMo Guardrails configuration.

Yes, it integrates with Presidio to detect and mask PII entities such as SSNs and email addresses in user messages. Define a mask pii subflow and enable the Presidio action parameter before processing.

False positives usually come from an overly sensitive detection threshold. Raise the jailbreak score cutoff (for example from 0.5 to 0.8) in your flow, and parallelize checks to keep latency low while tuning accuracy.

It runs on CPU, but a GPU such as an NVIDIA T4 with 4-8GB VRAM is recommended when integrating LlamaGuard. Typical overhead is 100-500ms depending on how many LLM-based checks are enabled.