add Llama Guard to your RAG pipelines to reasonable LLM inputs and outputs and fight immediate injection

LLM safety is an space that everyone knows deserves ample consideration. Organizations desperate to undertake Generative AI, from massive to small, face an enormous problem in securing their LLM apps. fight immediate injection, deal with insecure outputs, and stop delicate data disclosure are all urgent questions each AI architect and engineer must reply. Enterprise manufacturing grade LLM apps can’t survive within the wild with out strong options to deal with LLM safety.

Llama Guard, open-sourced by Meta on December seventh, 2023, affords a viable resolution to deal with the LLM input-output vulnerabilities and fight immediate injection. Llama Guard falls beneath the umbrella challenge Purple Llama, “that includes open belief and security instruments and evaluations meant to degree the taking part in discipline for builders to deploy generative AI fashions responsibly.”[1]

We explored the OWASP high 10 for LLM functions a month in the past. With Llama Guard, we now have a reasonably cheap resolution to begin addressing a few of these high 10 vulnerabilities, particularly:

• LLM01: Immediate injection
• LLM02: Insecure output dealing with
• LLM06: Delicate data disclosure

On this article, we are going to discover learn how to add Llama Guard to an RAG pipeline to:

• Reasonable the consumer inputs
• Reasonable the LLM outputs
• Experiment with customizing the out-of-the-box unsafe classes to tailor to your use case
• Fight immediate injection makes an attempt

Llama Guard “is a 7B parameter Llama 2-based input-output safeguard mannequin. It may be used to categorise content material in each LLM inputs (immediate classification) and LLM responses (response classification). It acts as an LLM: it generates textual content in its output that signifies whether or not a given immediate or response is secure/unsafe, and if unsafe primarily based on a coverage, it additionally lists the violating subcategories.”[2]