Building Reactive Large Language Model Pipelines with Motion

Large language models (LLMs) rely on prompts with detailed and informative context to produce high-quality responses at scale. One way to develop such prompts is through reactive LLM pipelines, which incorporate new information---e.g., end-user feedback and summaries of historical inputs and outputs---back into prompts to improve future response quality. We present Motion, a Python framework to build and execute reactive LLM pipelines. Motion uses a weak consistency model to maintain prompt versions, trading off freshness for end-user latency. We demonstrate Motion with an e-commerce application that suggests apparel to wear for any event, allowing attendees to indirectly influence prompts with their queries. Attendees can interact with the demo as end-users or modify the application as developers, adding new information sources for reactive prompts.