Jupybara: Operationalizing a Design Space for Actionable Data Analysis and Storytelling with LLMs

TL;DR

This work defines a design space for actionable EDA and storytelling built on semantic, rhetorical, and pragmatic dimensions, and demonstrates how to operationalize it with LLMs. It introduces Jupybara, a Jupyter Notebook extension that combines design-space-aware prompting with a multi-agent architecture to support exploratory data analysis and narrative generation. Through formative interviews and a user study with nine analysts, the authors show that Jupybara yields higher-quality, more steerable, and more explainable responses than web-based LLM tools, with the multi-agent mode outperforming the single-agent mode. The work argues that integrating EDA and storytelling within analysts’ workflows, along with explicit tracking of insights and transparent narration, can meaningfully enhance decision-making and communication in data-rich environments, while also offering a valuable testbed for human–AI collaboration in data science.

Abstract

Mining and conveying actionable insights from complex data is a key challenge of exploratory data analysis (EDA) and storytelling. To address this challenge, we present a design space for actionable EDA and storytelling. Synthesizing theory and expert interviews, we highlight how semantic precision, rhetorical persuasion, and pragmatic relevance underpin effective EDA and storytelling. We also show how this design space subsumes common challenges in actionable EDA and storytelling, such as identifying appropriate analytical strategies and leveraging relevant domain knowledge. Building on the potential of LLMs to generate coherent narratives with commonsense reasoning, we contribute Jupybara, an AI-enabled assistant for actionable EDA and storytelling implemented as a Jupyter Notebook extension. Jupybara employs two strategies -- design-space-aware prompting and multi-agent architectures -- to operationalize our design space. An expert evaluation confirms Jupybara's usability, steerability, explainability, and reparability, as well as the effectiveness of our strategies in operationalizing the design space framework with LLMs.

Figures (11)

• Figure 2: The multi-agent architecture for EDA in Jupybara. Given a user query, an Initial Respondent provides an initial response, which is then critiqued by four Critics. Based on the critiques, the Refiner improves the response and sends the revised version back to the Critics for review. The discussion between the Critics and the Refiner continues until all Critics agree the response is ready or a defined maximum number of discussion rounds is reached, at which point the final response is presented to the user. Notably, each dimension of the design space is addressed by at least three agents, potentially enhancing the quality of the response.
• Figure 3: The multi-agent architecture for data storytelling in Jupybara. An Initial Respondent generates a first draft of the data story, which is then critiqued by three Critics, each focusing on one dimension of the design space. The Refiner then improves the response and sends the revised version back to the Critics for further review. This process continues until all Critics agree the response is ready or the maximum number of discussion rounds specified is reached, at which point the final response is sent to the user. Notably, each dimension of the design space is addressed by three agents, potentially enhancing the quality of the response.
• Figure 4: (A) Rae inputs her command in a Notebook cell and clicks the "Invoke AI" icon in the cell toolbar. Jupybara inserts its response in a new Notebook cell and executes it, fulfilling the request. (B) Rae can configure the settings of Jupybara in the Settings tab. For each LLM agent in the system, Rae can choose between GPT-4o and Claude 3.5 Sonnet.
• Figure 5: Rae tasks Jupybara with a complex query under the single-agent mode. Utilizing its agentic workflow, Jupybara produces code, executes it, and provides an interpretation. Nonetheless, this response is not immediately intuitive to Rae due to the lack of units, the limited coverage of only 10 industries, and a textual format that is difficult to parse. (The side panel is hidden for clearer presentation.)
• Figure 6: (A) Rae switches to the multi-agent mode for the previous complex query. (See the video walkthrough in the supplemental materials for how to configure this in the Settings tab.) This time, Jupybara begins with an analysis plan in a markdown cell, executes the plan, produces a visualization, and finally generates an interpretation. Due to space constraints, only the code and the visualization are shown. Rae much prefers this response over the one in Figure \ref{['fig:eda single agent']}. (B) Rae wants to understand the rationale for Jupybara's treatment of null values and alternative strategies. She engages in a threaded conversation with Jupybara in the Clarify tab of the side panel.