Content area
Software development depends on Application Programming Interfaces (APIs), yet developers still struggle with incomplete, fragmented, or hard-to-navigate API documentation. Modern dialogue systems can provide conversational access to API knowledge in the form of chatbots and AI assistants, but building useful task-oriented assistants still demands clarity about how conversations should unfold, what capabilities a system should expose, and which data and tools can help enable them.
This dissertation provides the foundation for dialogue management in on-demand API documentation. First, it describes a "Wizard of Oz" study that yields a corpus of programmer-assistant interactions, as well as a multi-dimensional analysis that characterizes goals, dialogue acts, and grounding behaviors. Next, it develops a conversational dialogue manager optimized for efficient API search, and a clarification module that generates targeted questions to reduce ambiguity and steer retrieval. Finally, it demonstrates data-efficient integration with LLMs by using the dialogue manager to synthesize API search conversations, then fine-tuning a smaller model to exhibit desired behaviors (e.g., when to recommend a function versus ask for additional detail). Together, these contributions form a framework to engineer developer assistants whose behavior is grounded in empirical data.