Onboard custom real time detector on watsonx.ai+gov

A user should be able to configure a custom guardrail / real time detector on a Prompt Template (PTA) deployment. 

Context

A client does generative AI and enforce governance. As a consequence:

• The only way to consume a model in Production is through a deployed Prompt Template (PTA)
• Prompt Template is tracked in an AI use case
• Prompt Template has been evaluated with multiple tests (OOB metrics from x.gov, custom metrics, LLM as a Judge, Human validation)
• Monitoring is activated on the Prompt Template

Requirements

Now that the prompt template is in Production we want to add an additional line of defence by executing tests (aka. Guardrails / real-time detector) in real-time on the PTA input and/or output (priority is on output). The outcome of the test could be one of the following:

Blocking behavior

• If the test if OK or with a score above a certain threshold then the output is sent back to the consuming application
• If the test KO or below a certain threshold then the output is blocked and an error is sent back to the consuming application

Non-blocking behaviour --> priority

• The ouput of the test is described in json format and appended to the body of the response (like PII/HAP moderation today). Then it is the responsability of the consuming application to decide what to do.

Since a guardrail can be code, regex, SLM, or LLM-based, the most simple way would be to be able to deploy guardrail as Python function with a standardized interface contract. Then, we should execute some configuration steps to apply the guardrail on a particular deployed PTA. The end to end flow to build and deploy a PTA would be as follows:

1. Experiment on the PTA
2. Create a new test to detect something on a PTA (e.g. build a model that detect if the outcome of the model respect the brand tone & voice)
3. Use evaluation data to validate that the test is performant enough
4. Deploy the test as Python function and make it a custom metrics
5. Use this custom metric in the evaluation phase to validate PTA
6. Track, evaluate again and deploy PTA
7. Configure PTA deployment to have test executed in real time and test outcome included in the response body --> here the test used during evaluation phase becomes a guardrail
8. Guardrail output must be logged in Openscale with payload data