Exploring NeMo Guardrails’ sensible use circumstances

On the subject of LLM safety, we’ve explored OWASP prime 10 for LLM purposes, Llama Guard, and Lighthouz AI so removed from completely different angles. At the moment, we’re going to discover NeMo Guardrails, an open-source toolkit developed by NVIDIA for simply including programmable guardrails to LLM-based conversational programs.

How is NeMo Guardrails completely different from Llama Guard, which we dived into in a earlier article? Let’s put them aspect by aspect and examine their options.

As we will see, Llama Guard and NeMo Guardrails are essentially completely different:

• Llama Guard is a big language mannequin, finetuned from Llama 2, and an input-output safeguard mannequin. It comes with six unsafe classes, and builders can customise these classes by including extra unsafe classes to tailor to their use circumstances for input-output moderation.
• NeMo Guardrails is a way more complete LLM safety toolset, providing a broader set of programmable guardrails to manage and information LLM inputs and outputs, together with content material moderation, matter steering, which steers conversations in the direction of particular matters, hallucination prevention, which reduces the technology of factually incorrect or nonsensical content material, and response shaping.

Let’s dive into the implementation particulars on tips on how to add NeMo Guardrails to an RAG pipeline constructed with RecursiveRetrieverSmallToBigPack, a sophisticated retrieval pack from LlamaIndex. How does this pack work? It takes our doc and breaks it down, beginning with the bigger sections (mum or dad chunks) and chopping them up into smaller items (youngster chunks). It hyperlinks every youngster chunk to its mum or dad…