Scoring models in Plexus follow a uniform interface standard based on the MLFlow pyfunc standard. This standard is designed to ensure that all scoring models in Plexus have a consistent interface, regardless of the model type or the specific implementation. Using a standard interface enables mixing and matching different types of classifiers in one scorecard.
MLFlow provides a unified approach for creating and managing custom MLflow models. This interface standard allows any model to be integrated into the MLFlow model registry using mlflow.pyfunc.log_model(). At runtime, these models can be retrieved with mlflow.pyfunc.load_model() for inference. This standard interface ensures consistency in sending transcript text to the model and receiving responses, whether dealing with a machine learning model, an agentic LLM score, or a programmatic model, as demonstrated in the first example of the provided guide. For more detailed information, refer to the documentation for mlflow.pyfunc.
The specific details of the standard are:
Every Score in Plexus derives from mlflow.pyfunc.PythonModel and must implement a predict() method.
This method must accept a Input object instance. The Input classis defined in the plexus.MLClassifier module.
The predict() method must return a Result. The Result is defined in the plexus.MLClassifier module, and each classifier can optionally define additional Result subclasses to support additional features.