Resilience policies evaluate your data flows and apply rules based on defined topics. They can block or issue warnings for inputs, outputs, or both.
The goal is to enable the implementation of data policies to ensure a certain level of quality in AI responses and the topics the AI is permitted to address. This mechanism is based on the “LLM as judge” approach, in which another AI analyzes what has been said and issues a judgment validating or invalidating the content.
Configure a policy
You configure a policy on the Resilience Center screen. Start by giving your new policy a name (and a description). You will then have a number of settings available.
Target
The target defines which data the policy will apply to. By default, it will apply to all API connectors (chatbots) as well as your organization’s or workspace’s Studio. For connectors, you can refine your selection by clicking an “Advanced” link.
Flux
The flow determines when the policy is evaluated. The options are:
- Input: The AI analyzes what a user has written
- Output: The AI analyzes what the system has generated
- Bidirectional: A combination of the two previous options
Taxonomies
A taxonomy is a set of rules that tells the system how to classify and evaluate responses. Our platform offers two types of taxonomies: predefined and custom.
Predefined taxonomies are the standard options and allow you to enable these protections with a single click. Currently, they include topics such as violence/harm, sexual content, personally identifiable information, and more.
Custom Taxonomy
Custom taxonomies allow you to go a step further and address specific needs related to your business. The parameters to provide are:
- Name: Used by the AI to refer to the taxonomy, so it must be concise and descriptive, and written in English
- Definition: Used by the AI to understand what content falls under the taxonomy; it must therefore be detailed and precise regarding the content it covers. It must be written in English.
- Warning message: The text that will be displayed to the user if this taxonomy is detected.
Exemple avec le blanchiment d’argent:
Name: Money Laundering / Financial Crime
Definition: Content that provides guidance, advice, or instructions related to money laundering, illegal financial transactions, or circumvention of financial regulations.
Warning message: I’m unable to provide guidance on illegal financial activities, including money laundering.
Action
Actions describe what is executed when a resilience policy is detected. Either the system will notify (the user) of the detection, or the system will reject the current action. This second action ensures, for example, that the system never responds with actions that would be unethical.
Other
When saving, you can choose to enable or disable a policy, and enable it at a later time.
It’s also worth noting that it’s possible to have multiple policies on the same API connector or Studio.