Generative Data Visualization with JSON Representations

Data visualization can provide a wealth of insights for any organization. However, creating data visualizations can take significant time and domain-specific knowledge. Generative data visualization has the potential to streamline this process by creating complex and insightful visualizations with minimal user intervention. Large Language Models (LLMs) can respond to natural language queries with various media. However, they have yet to be widely applied to data visualization tasks in an enterprise setting due to the complexities of model training, resource consumption, and upkeep. This study tests an approach to generative data visualization through in-context learning, using an LLM to output user data visualization requests in the JSON format. Overall, the LLM performs poorly, even with multiple in-context examples. However, the study also yields insights into targeted adjustments that can potentially boost in-context learning for this use-case.