Guardrails for safe implementations of AI-based services

The creation of generative AI enabled services requires multiple steps, including data acquisition, data management, training, fine-Tuning of models, or generation of prompts and documents for services such as those that can be used in a RAG (Retrieval Augmented Generation) process. The quality and characteristics of data ingested and generated during these stages can have a tremendous impact on the characteristics of the service that is produced. In order to have a safe and assured operations, safety guard-rails for data have to be introduced at each stage of the process. In this paper, we present an overall architecture for guard-rail management, provide a taxonomy of guard-rails, and discuss the characteristics of the guard-rails that will be required for operations at each stage of the process.