This paper focuses on the transition of software development toward AI-integrated applications that leverage generative AI and large-scale language models (LLMs) at runtime. While existing LLM integrations introduce complexity due to the difficulty of prompt engineering, our proposed Semantic Type Programming (MTP) addresses this challenge by abstracting LLM integration through intuitive language-level constructs. MTP leverages the semantic richness of code to automate prompt generation and response processing without additional developer effort. Its core components include the by operator for LLM invocation, a semantic-based intermediate representation (MT-IR), and MT-Runtime, an automated system for managing LLM interactions. We implement MTP in the Jac programming language, a superset of Python, demonstrating that it significantly reduces code complexity while maintaining correctness and efficiency. User studies show that developers using MTP work 3.2x faster and with 45% fewer lines of code compared to existing frameworks. Furthermore, we demonstrate robustness even when naming conventions are degraded by up to 50%. MTP is developed as part of the Jaseci open source project and is available in the byLLM module.
