Evolving AI-Driven Workflow Management, Part A: Strategies for Token Window Challenges and Utilization of Provenance Data

Abstract

Product development in technical applications has become a highly complex process and is increasingly supported by sophisticated software systems. Traditional workflow management environments for automating the required processes have become very helpful tools. However, the complexity of these expert systems poses significant challenges to engineers and requires increasing levels of expertise. In the past, various approaches provided support to the user in generating complex workflows. Large Language Models (LLMs) as part of the natural language processing have great potential as assistance systems for centralizing expert knowledge. Part A of this two-part paper extends an existing method for automating workflow generation. The focus addresses the challenge of the limited context window length of LLMs. Several approaches have been analyzed and investigated. In addition, a provenance data management system is integrated so that historical information can be included in the generation using LLMs. Part B addresses the challenge of dealing with several possible workflows or ambiguous workflow solutions.