ReferenceAuthoring Portal Reference
Variables
The Variables Section is used to manage the set of defined variables used in your Agent.
A Variable must be defined here before it can be updated or exposed to the LLM in any Stage.
Adding a Variable Definition
-
Click the ‘+’ button at the top-right of the section
-
A new blank variable will be added to the list of defined Variables
Editing A Variable Definition
Fill out the configuration for the Variable
| Field | Details |
|---|---|
| Name | A descriptive name for the variable. Use an informative name as this is exposed to the LLM and its value conveys information |
| Type | The type of the variable. Select the most appropriate type. This type will be used by the LLM model when extracting data from the conversation and will also be used when calling integrations. |
| Purpose | Text explicitly describing how this variable is being used. This is exposed to the LLM and is the main mechanism for ensuring that the LLM understands what information you wish to extract and when to extract it or modify it. |
| Default Value | An optional value to use if the LLM has not set the value yet or it has cleared the value. |
| Advanced | |
| Externally Initialized Variable | Specifies whether this value is expected to be provided by a front-end when a session involving this Agent is initialised. To mark a variable as externally initialised means you are intending to initialise it and set its value when a session is started by an external front end. Some examples where you might use this feature include passing in the user’s name, region, or anything else specific to the session that you want to use in the conversation and don’t need to collect from the user because your external system already knows the value. If you specify that a variable as an Externally Initialized Variable and fail to provide a value when initialising a session, your Agent will generate an error and fail to run. To support Sandbox testing of Agents using this type of variable, you can specify the initial value in a sub-section that will appear at the bottom of this section if required |
| Substitute Value | Use this field to provide a substitute value that will be shown to the LLM instead of the actual value. This can be used to ‘mask’ variables that might enable the LLM to display bias e.g. variables that expose gender or private data (Personally Identifiable Information). For example, if you use a substitute value for a value holding the user’s name (to replace it with a gender neutral value), the LLM will only see that gender neutral value, preventing it from inferring the user’s gender from their real name. The inference engine replaces any references that the LLM inserts into the response with the real value so that the end user’s experience still feels personalised. Due to the pattern matching required when rolling back this value if it has been used by the LLM in AI responses, you must only provide a single word value. |
Deleting A Variable Definition
Click the trash icon for the variable