The gap between non-technical users and tabular data limits efective access and data-driven decision-making. To address this challenge, this work presents an agent-based workflow integrating text to visualization (Text-to-VIS) and data storytelling using Large Language Models (LLMs). Given a natural language query and a dataset, the system generates Python code to produce charts and provides narrative insights to support interpretation. Using GPT-4o, the approach is illustrated through three proof-of-concept demonstrations on a movie dataset. The results show that LLMs can generate appropriate visualizations and assist users in understanding patterns and key findings, bridging the gap between data visualization and storytelling.
Data visualization is vital in enabling users to understand complex data through graphical
representations. Traditionally, creating efective visualizations required proficiency in data analysis and
programming. Recently, the advent of Large Language Models (LLMs) such as ChatGPT1 and the
exploration of prompt engineering techniques have opened new avenues for developing more efective
and user-friendly natural language interfaces (NLI) for data interaction [
While Text-to-VIS systems focus on generating visual representations, they often lack the narrative
element that helps users interpret and communicate the insights derived from data. This is where data
storytelling becomes essential. More than just displaying data clearly and eficiently, data storytelling
is a methodology for conveying key insights to a specific audience to inform decisions and guide
action. It involves structuring the analytical message as a narrative that enhances the communication
and dissemination of findings within and across groups [
ISSN1613-0073
This work explores the integration of Text-to-VIS and data storytelling using LLMs in a unified workflow. It presents a workflow that combines chart generation with automated narrative generation, powered by GPT-4o, to support intuitive, insightful, and engaging data exploration. Through a series of exploratory demonstrations, the work illustrates how LLMs can be used to construct relevant visualizations and generate compelling explanations tailored to the user’s analytical goals.
Text-to-VIS is the object of growing attention with the rise of LLMs. Early work, such as Data2Vis [
In parallel with advances in Text-to-VIS, several works have focused on integrating data storytelling
into visualization workflows to enhance interpretability and engagement. “Calliope” [
The system is implemented using LangChain3 and LangGraph4, which enable the creation of modular and agentic workflows based on LLMs. As illustrated in Figure 1, the system is composed of four primary components: Clarify Question, Human Feedback, Text-to-VIS, and Data Storyteller. Each component is modeled as a node within a directed workflow graph, where predefined paths govern transitions but are conditionally triggered based on the output of LLM decisions.
• Clarify Question: The system first evaluates the clarity of the user’s natural language query
using an LLM, following the LLM-as-a-Judge paradigm [
The evaluation of the proposed system is based on three proof-of-concept demonstrations, using the
GPT-4o model provided by OpenAI [
The first demonstration explores a question that analyzes how LLMs handle time-based data requests. The model generated a bar chart (see Figure 2(a)) showing the annual number of movie releases from 1996 to 2010. The narrative generated by the LLM provided a title (“The Evolution of Movie Releases Over Time”), contextualized the trend, and identified key insights: Growth (1996–2000), Stability (2001–2007), and Sharp Decline (2008–2010). It framed the decline as a conflict ( “A Noticeable Decline in Recent Years”) and proposed causes, such as economic crisis, technological shifts, and industry saturation. (a) Result from Demonstration 1. (b) Result from Demonstration 2.
The second case illustrative example explores how LLMs generate visual summaries from categorical and aggregated numerical data, focusing on average gross income grouped by movie genre. The Textto-VIS Tool produced a bar chart ranking genres and autonomously sorted the chart by average gross 5https://github.com/nl4dv/nl4dv/blob/master/examples/assets/data/movies-w-year.csv income (without explicit prompting) as shown in Figure 2(b), revealing Adventure as the top-performing genre. Its narrative, titled “The Power of Genre in Box Ofice Success” , framed the revenue disparity as a conflict ( “The Uneven Playing Field of Genres”) and highlighted key insights: Adventure’s dominance (blending action/fantasy), strong performance of Action/Musicals, and niche genres’ struggles. Notably, the LLM dedicated a paragraph to hypothesize Why adventure genre reigns Supreme, citing broad demographic appeal and high production value. Thus, this demonstrates the model’s capability to generate interpretive analysis.
The third demonstration analyzed budget-gross correlation and began with Clarify Question, asking for more information: “Do you want to analyze the correlation for all movies or by specific genre/period?” . When prompted specifically for adventure genre, the Text-to-Vis tool was called and generated a scatter plot (see Figure 3) showing a positive but non-linear relationship. The Data Storyteller emphasized that, while higher budgets generally yield higher grosses, exceptions exist, such as low-budget hits and high-budget flops. Finally, it concluded that captivating storytelling ultimately outweighs pure ifnancial investment.
The explorations suggest that LLMs can generate accurate visualizations and meaningful narratives supporting user interpretation. The charts matched the queries and data, while the narratives followed a coherent arc: providing context, highlighting trends, and suggesting explanations.
The LLM-based question clarity check was efective in prompting clarifications only when needed. Still, the quality of the narratives depends on prompt design and the model’s interpretation, which may lead to generic insights. Overall, the workflow helps bridge the gap between visualization and understanding, and shows potential for building more intuitive tools for non-technical users.
All prompts, LLM outputs, generated charts, and full narratives from these experiments are available in a public GitHub repository6.
This work presented a workflow that integrates Text-to-VIS and data storytelling using LLMs. Proof-ofconcept demonstrations based on the Movies dataset demonstrated the system’s ability to transform natural language queries into relevant charts and provide insightful narrative explanations with minimal 6https://github.com/dudursn/llm_based_text2vis_and_storytelling user intervention. Future work will explore multiple Text-to-VIS strategies and incorporate an agentbased evaluation module to select the best visualization alternative. It is also planned to assess the quality of the generated story and its visualization, with a particular focus on personalization. One key challenge is leveraging user intent to guide the storytelling process in a meaningful and adaptive way.
During the preparation of this work, the authors used GPT-4o and Grammarly for grammar and spelling checks, and take full responsibility for the final content.
All prompts, LLM outputs, generated charts, and full narratives from these experiments are available via • GitHub at https://github.com/dudursn/llm_based_text2vis_and_storytelling