Create a custom guardian

Create your own custom guardian to monitor and detect requests sent to LLM.

Before you begin

Configure a connection and credential alias for your preferred provider. For more information, see Create a Connection & Credential alias.

Role required: admin

Procedure

1. Navigate to All, and then enter sys_generative_ai_custom_guardian_transformer_list.do in the filter to go to the Generative AI Custom Guardian Transformer [sys_generative_ai_custom_guardian_transformer_list.do] table.
2. Select New.
3. In the Name field, enter the name of the custom guardian.
4. Create a request transformer script.
   1. In the Request Transformer filed, enable the Turn on ECMAScript 2021 (ES12) mode toggle to create script.
      The following example shows the request transformer script for the custom guardian request structure.

      ((function(inputs) {     
      	/*        
      	inputs structure: {     
      		prompt: object, The textual content that needs to be evaluated with guardian           
      		threshold_setting: object           
      		connection_attributes: object        
      	}     
      	write code here to construct the request body and any custom headers needed using the inputs object.     
      	*/     
      	// construct body using the inputs like input.prompt and so on.      
      	var body = {};      
      	//construct headers using the inputs      
      	var headers = {};      
      	return {     
      		body: body,     
      		headers: headers     
      	};     
      })(inputs);
      
      	 

5. Select the Active option.
6. Create a response transformer script.
   1. In the Response Transformer filed, enable the Turn on ECMAScript 2021 (ES12) mode toggle to create script.
      The following example shows the response transformer script for the custom guardian response structure.

      (function(inputs) {     
      	/*     
      	write code here to transform the custom guardian response into standard guardian response format expected by platform     
      	inputs structure: {           
      		response_body: object,     
      		threshold_setting: object     
      	}     
      	*/     
      	// write code here to populate the response object.     
      	var guardianResponse = {};     
      	//Set flagged to true, if the guardian has flagged the text. Categories are optional and are for logging purpose only     
      	// if transformer setting is Prompt Injection, expected response format: {security:{flagged:false,categories:[]}}     
      	// if transformer setting is Offensiveness, expected response format: {safety:{flagged:false,categories:[]}}     
      	// if transformer setting is ALL, expected response format: {safety:{flagged:false,categories:[]},security:{flagged:false,categories:[]}}      
      	return guardianResponse;     
      })(inputs);
      
      	 

7. In the Guardian Setting field, select a guardrail of your choice.
   • Select All to handle the detection of both offensiveness and prompt injection together.
   • Select Offensiveness to enable the detection of offensive or harmful content.
   • Select Prompt Injection to enable the detection of prompt attacks.
   Note:

   If your LLM can evaluate both offensiveness and prompt injection in a single request, select All. If not, create separate guardrail records for each offensiveness and prompt injection.
8. In the Connection And Credential Alias field, select the alias that you want to integrate with your custom guardian.
9. Select Submit.
10. Set up the custom guardian token limit.
    1. Navigate to All > System Definition > Tables.
    2. Select the Gen AI Control Setting Data [sys_gen_ai_control_setting_data] table.
    3. In the Related Links section, select Show List.
    4. Search for the Custom Guardian Token limit record in the Gen AI Control Data column.
    5. Update the Value field of the Custom Guardian Token limit record with the maximum number of tokens per request.
       The token limit defines the maximum number of tokens the guardrail can process in a single evaluation. If the request content exceeds this limit, the content splits into smaller parts. Each part is evaluated separately against the guardrail rules.

       For example, if the token limit is 1000 and the request content is 2500 tokens, the system splits it into three parts and evaluates each one individually.