Validate AI Models and Applications

• Currently, only AWS Bedrock is supported. You must configure the AWS Bedrock integration. See the section, AWS Bedrock Integration.

• A Multicloud Defense integration is required for model discovery. See Multicloud Defense Integration.

A model validation test requires a set of parameters that will be used to connect to and test the AI model. To configure and run a validation test for a model:

1. Click the Run validation button in the Validation page.

2. In Asset type, choose Model.

3. Specify the test parameters and the model to be tested:

   • Asset name: The name of the model as shown in the Assets page.

   • Model ID: The model ID as stored in the platform hosting the model.

   • Test name: Give this test a memorable name to better find it later.

   • Prompt template: This is the JSON request payload that will be sent to the model's inference API in order to test it. This must include a {{prompt}} placeholder where the AI Defense-generated test prompts will be inserted.

   • Response: a JSON path that specifies where in the HTTP response. See “Format the response path” below.

4. Click Submit. The test runs immediately.

• Create the application and connection in the AI Defense: Applications panel.

• Have your model API credentials ready for including in the HTTP headers used to connect to the model.

The test requires a set of parameters that will be used to connect to and test the application’s AI model. Follow the steps below to configure and run a validation test for an application.

An application validation evaluates the model connected to the application, not the application itself.

1. Click the Run validation button in the Validation page.

2. In Asset type, choose Application.

3. Specify the test parameters and the model to be tested:

   • Application: Application to be tested. This is the name you gave to the application in the AI Defense: Applications panel.

   • Test name: Give this test a memorable name to better find it later.

   • Endpoint: The endpoint of the LLM used by the application

   • Inference API path: The API path for model inference calls. For example: /openai/deployments/gpt3.5

   • Prompt template: This is the JSON request payload that will be sent to the model's inference API in order to test it. This must include a {{prompt}} placeholder where the AI Defense-generated test prompts will be inserted.

   • Response: a JSON path that specifies where in the HTTP response. See “Format the response path” below.

   • HTTP headers: Headers for the inference API connection. Specify the authorization values here.

4. Click Submit. The test runs immediately.

Have your model API credentials ready for including in the HTTP headers used to connect to the model

Follow the steps below to configure and run a validation test for an external model.

1. Click the Run validation button in the Validation page.

2. In Asset type, choose External.

3. Specify the test parameters and the model to be tested:

   • Test name: Give this test a memorable name to better find it later.

   • Target: The fully qualified URL of the external model endpoint

   • API provider: Optional. Pick the API provider, such as OpenAI, to populate the request and response templates.

   • Request body template: This is the JSON request payload that will be sent to the model's inference API in order to test it. This must include a {{prompt}} placeholder where the AI Defense-generated test prompts will be inserted.

   • Path to model response: a JSON path that specifies where in the HTTP response. See “Format the response path” below.

   • HTTP headers: Headers for the inference API connection. Specify the authorization values here.

4. Click Submit. The test runs immediately.

To retrieve a response from a nested JSON field, use dot notation to specify the levels of nesting. For example, if the endpoint returns a nested response like the one shown below, then you would set your Path to model response to response.llmResponse. This works for a response format like:I am an AI Chatbot, how can I assist you?

To extract a response string from an array, specify the element's index in square brackets. For example, if the endpoint returns a nested response like the one shown below, you would set your Path to model response to content.[0].text. This works for a response format like:

{
    "content": [
        {
            "text": "Bonjour, je suis Claude!",
            "type": "text"
        }
    ],
    "id": "msg459674598",
    "model": "claude-3-5-sonnet-2024-08-20",
    "role": "assistant"
}