Introduction to Guardrails

Introduction

Guardrails filter out ill-intentioned and malicious, ensuring safe interactions between users and language models.

As organizations deploy language models for public use, implementing Guardrails is crucial to protect users and maintain trusts in AI systems.

This technical will explain how Guardrails works and their real-world benefits.

Why Guardrails?

Safety: Guardrails ensures user prompts and model responses are safe. There are some cases users abuse language models to spit out sensitive informations. Also, there are extreme cases where language models forced or coaxed underage children to do the unthinkable.

Cost Efficiency: By filtering malicious prompts before they reach language models, we prevent wasted GPU cycles and reduce operation expenses.

Auditability: Guardrails enable better system auditing and compliance tracking, making it easier to identify abuse patterns and enforce usage policies.

What are Guardrails?

Guardrails are a set of rules, filters, or mechanisms that ensures content or prompts adhere to certain safety, ethical, or legal standards.

What are Guardrails filtering out?

1. Biased and harmful expressions
2. Abusive and unethical behaviours
3. Illegal and malicious activities

Simply said, you can imagine Guardrails as a bouncer at a club, checking the guest list and only allowing in those who meet the criteria for entry. However, Guardrails do more than that. They are like a security team that anticipates risks, analyze contexts, and adapts to new threats.

How Guardrails work?

Each Guardrails system could have many layers of filtration, the following are three common ones:

1. Keyword filtering
2. Regex filtering
3. Semantic filtering

Keyword Filtering

With this technique, one could keep a list of words that do not align with one’s business or ethical standards. This is the easiest and most straightforward way to implement Guardrails.

Regex Filtering

Regex filtering is a more advanced technique that allows for pattern matching. This can be useful for filtering out ID numbers, email addresses, phone numbers, or even credit card numbers.

Semantic Filtering

Semantic filtering uses language models to understand the real meaning behind each prompt, both its context and intent. This is also one of the easiest technique to incorporate into Guardrails.

Challenges of Guardrails

Keywords filters

This filtration is also the least effective since languages are considered “alive” as long as people are using them. People would try to use new keywords or misspell existing keywords to bypass the filters. Thus the keyword list needs to be updated frequently to keep up with the evolving language.

Regex filters

A list of regex can be highly complex and difficult to maintain, especially as new patterns emerge. As the list of regex rules grows, it can also lead to CPU overhead and increased false positives, where legitimate prompts are mistakenly blocked.

Semantic filters

With this filtration, one should be ready to face with the monthly bill from the language model provider. In addition to that, frequently accessing LLM APIs would hit the rate limit and cause latency issues.