Jupybara : Operationalizing a Design Space for Actionable Data Analysis and Storytelling with LLMs Conference: CHI Conference on Human Factors in Computing Systems; April 26-May 1, 2025; Yokohama, Japan CHI Conference on Human Factors in Computing Systems (CHI ’25), April 26-May 1, 2025, Yokohama, J
Huichen Will Wang, Larry Birnbaum, Vidya Setlur · 2025
Read Jupybara : Operationalizing a Design Space for Actionable Data Analysis and Storytelling with LLMs Conference: CHI Conference on Human Factors in Computing Systems; April 26-May 1, 2025; Yokohama, Japan CHI Conference on Human Factors in Computing Systems (CHI ’25), April 26-May 1, 2025, Yokohama, J by Huichen Will Wang, Larry Birnbaum, Vidya Setlur (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-16661.txt.
Introduction
Huichen Will Wang, Larry Birnbaum, and Vidya Setlur present Jupybara, an AI helper for data work, in a twenty twenty-five paper. The hard task they target is exploratory data analysis: digging in complex data and telling a clear story about what matters. Good stories need exact facts, strong framing, and clear relevance.
The team first maps a design space for such work, drawn from theory and talks with experts. Jupybara, built as an add-on for Jupyter Notebook, turns that map into prompts and a team of agents. Expert tests found it easy to use, steer, and fix, and found the design map a sound base for the tool.
Catalog record
| CREATOR | Huichen Will Wang, Larry Birnbaum, Vidya Setlur |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 99 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-16661
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Jupybara : Operationalizing a Design Space for Actionable Data Analysis and Storytelling with LLMs Conference: CHI Conference on Human Factors in Computing Systems; April 26-May 1, 2025; Yokohama, Japan CHI Conference on Human Factors in Computing Systems (CHI ’25), April 26-May 1, 2025, Yokohama, Japan DOI: 10.1145/3706598.3713913 ISBN: 979-8-4007-1394-1/25/04 CCS: Human-centered computing Visualization systems and tools CCS: Human-centered computing Visualization theory, concepts and paradigms Huichen Will Wang, Larry Birnbaum, Vidya Setlur Abstract . Figure 1. The interface of Jupybara , an AI-enabled assistant for actionable EDA and data storytelling implemented as a Jupyter Notebook extension. Jupybara operationalizes our proposed design space consisting of the semantic , rhetorical , and pragmatic dimensions. (A) For a complex user query in EDA, Jupybara identifies and presents an analysis plan before producing code. (B) In the data story generated by Jupybara , the system uses precise language to convey analytical results; appropriate hooks, connectives, and narration of analytical strategies to bolster actionable insights; and relevant domain knowledge to connect data facts to actionable insights. 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. † † cc-license: by 1. Introduction Ben Schneiderman famously noted, “the purpose of visualization is insight, not pictures.” ( Hullman 2019 ) Similarly, the goals of data analysis and storytelling extend beyond merely generating statistical output, data visualizations, or even data narratives: these activities are fundamentally about extracting and communicating insights. In this work, we focus on a particular class of insights, actionable insights. Traditionally, researchers have viewed insights as a collection of knowledge over the data ( Battle and Ottley 2023 ; Chang et al. 2009 ) , be it data facts, utterances, or knowledge links. A defining hallmark of actionable insights is that they not only sit on knowledge over the data but also aim to empower people to make informed decisions and take sensible actions, thus often requiring strategies tailored to real-world questions to derive. The decision- and action-oriented nature of these insights makes them challenging to generate but crucial for solving practical problems. In practice, many analytic processes are directed toward generating actionable insights. By combining data analytics with domain knowledge, these processes bridge the gap between raw data and strategic actions. For example, the entire practice of business analytics centers on transforming data into actionable insights that drive competitive advantage and organizational success by leveraging industry know-how ( Delen and Ram 2018 ) . Educational data mining synthesizes educational psychology with data mining techniques to resolve challenges in learning ( Romero and Ventura 2010 ) . An important branch of epidemiology is concerned with analyzing data to propose interventions for the spread of diseases that are grounded in biological and clinical theories ( Brachman 1996 ) . In this work, we interviewed nine expert data analysts to map a design space for exploratory data analysis and storytelling for actionable insights (hereafter actionable EDA and storytelling 1 1 1 We define actionable EDA as the analytical process of deriving actionable insights, and actionable data storytelling as the process of crafting a data story to convey actionable insights to an audience. ) and considered them in foundational theories in data science, narrative discourse, and communication. We illustrate how semantic precision (accurately interpreting, tracking, and verbalizing EDA findings), rhetorical persuasion (strategically mining and communicating results to corroborate actionable insights), and pragmatic relevance (integrating EDA findings with domain expertise) collectively contribute to sound and effective actionable EDA and storytelling. The interviews also helped us characterize the workflows of actionable EDA and storytelling (fluid, integrated, and iterative) and uncover key challenges in the process, such as identifying appropriate analytical strategies and leveraging relevant domain knowledge. We then show how our proposed design space can inform the development of better AI assistants for actionable EDA and storytelling. While these tasks have traditionally relied on human expertise, Large Language Models (LLMs) present exciting opportunities to augment human expertise with machine intelligence. With their statistical knowledge, LLMs can amplify human capabilities in extracting meaningful patterns from data. Furthermore, trained on vast corpora, LLMs can potentially generate coherent narratives and apply domain expertise to interpret analytical results, enabling analysts to make more informed strategic decisions. We propose two strategies to utilize LLMs in operationalizing the framework: design-space-aware prompting (prompting LLMs with guidance distilled from the design space) and multi-agent architectures (improving response quality along the dimensions of the design space through multi-agent collaboration). We further implement these strategies in a Jupyter Notebook plugin supporting actionable EDA and storytelling, Jupybara 2 2 2 “ Jupybara ” is a portmanteau of “Jupyter” (Notebook) and “capybara,” a friendly and likable creature that satisfies Gricean Maxims of cooperative behavior. Known for their social nature and tendency to stack on top of each other, capybaras symbolize the multi-agent architecture of the tool for facilitating actionable EDA and storytelling. (Figure 1 ). In addition to being an agentic EDA copilot and data storyteller, Jupybara allows users to seek clarification, track insights, and edit AI-generated content manually or with AI assistance. A preliminary evaluation with nine data analysts showed that participants found Jupybara produced higher-quality responses for actionable EDA and storytelling compared to web-based LLM tools (e.g., ChatGPT). Participants also rated Jupybara as more usable, steerable, explainable, and reparable. Moreover, participants deemed the multi-agent mode of Jupybara to produce better responses than the single-agent mode across the three dimensions of our design space. In sum, our primary contributions are: • A design space that integrates the semantic , rhetorical , and pragmatic dimensions to facilitate effective actionable EDA and storytelling; • A formative interview and participatory design session with nine data analysts that identified analysts’ workflows, considerations, and challenges in actionable EDA and storytelling; • Two strategies to operationalize the design space, design-space-aware prompting and multi-agent architectures, which we implement in a Jupyter Notebook plugin supporting actionable EDA and storytelling, Jupybara ; • A user evaluation of Jupybara with nine participants demonstrating the usability, steerability, explainability, and reparability of Jupybara and the effectiveness of our strategies for operationalizing the design space. 2. Related Work Our work is informed by research in the use of language in visual analytics and human-AI collaboration for data analysis and storytelling. 2.1. The Use of Language in Visual Analytics The integration of natural language processing (NLP) techniques with visual analytics tools has explored how language can be used to express user intent, bridging analytical inquiry