SemPiper Synthesizes ML Pipeline Code with Semantic Operators

Developing machine learning (ML) pipelines is often a complex and error-prone process, largely due to the extensive data preparation, feature engineering, and integration required across diverse data sources. While large language models (LLMs) have shown potential in assisting programming tasks, current chat-based interfaces offer limited control over pipeline behavior and often generate code that is difficult to optimize or integrate into production systems. SemPipes addresses these challenges by proposing a new programming model that enhances ML pipelines with declarative, LLM-powered semantic data operators. This allows developers to specify high-level natural language instructions for data-centric operations, which can then be seamlessly combined with conventional Python code from established data science libraries. The system synthesizes specialized implementations for these semantic operators during pipeline training, adapting them based on dataset characteristics and the overall pipeline context. This approach enables flexible yet controlled integration of LLM capabilities into ML development. The demonstration, SemPiper, provides an interactive interface that visualizes the computational graphs of these pipelines, the synthesized operator implementations, and the optimization trajectories derived from an evolutionary search procedure. Users can explore various end-to-end scenarios, modify pipelines, inspect the generated code, and observe the synthesis and iterative optimization of semantic operators. This highlights how declarative semantic operators can lead to more controllable, optimizable, and practical integration of LLMs into the ML pipeline development workflow.