watsonx.ai invocation from ODM

(more background in this Medium post)

This repo provides a Java XOM library (see watsonx-helper Maven project) to support the invocation of a watsonx.ai inference from ODM rules.

The WatsonxAIRunner.java class allows to:

• Store the different parameters for the LLM invocation (prompt template, prompt variables and model parameters). Most of the model parameters are associated with a default value.
• Run a watsonx.ai inference using the above parameters, the watsonx.ai instance URL and the API key for this instance.

The class is verbalized in the watsonx-model BOM entry in the rule project, allowing to:

• Define the prompt template to be executed and the prompt variables. Below is an example of parameters setup:

  

• Execute the inference on the parameters:

  

Simple use case

The simple use case for testing is about deciding how to dispose from a customer review, either publish it if the review is positive or ignore it if it's negative.

• The ODM input provides the customer review text.
• A simple sentiment extraction prompt template return either positive or negative as a result.
• ODM passes the prompt template instantiated with the review text to the LLM.
• The LLM response is mapped to an enumeration and used in a decision table to decide how to dispose of the review.

Sample execution

• Run mvn install package in the watsonx-helper folder to compile and package the the uber jar used for the ODM XOM.
• Run the reviewDisposition decision operation to test the rules.