Making a chart involves more than selecting a chart type. First I need to understand the dataset, decide what question is worth asking, and choose a representation that answers it.
A short LIDA experiment in my files exposes those stages through a small script. It creates an OpenAI-backed Manager, summarizes data, proposes visualization goals, and requests a visualization.
Separating the steps
The summarize, goals, and visualize calls make the workflow easy to read. A dataset summary helps describe what the columns contain. Goal suggestions propose things a person might investigate. Visualization generation then attempts to turn a selected goal into chart code.
The distinction is useful because each stage can go wrong differently. A column's meaning may be misunderstood before any chart code is produced. A technically valid chart may also answer an unhelpful question.
The LIDA paper, first submitted in March 2023, describes this broader staged approach to visualization generation.
Reviewing the generated analysis
For an actual dataset, I would inspect the proposed goal before the chart. Does it make sense for the units, categories, and time range? Has the model assumed that a code or identifier is a numerical measure?
After generation, I would inspect the data transformation and rendered chart together. Sorting, aggregation, missing values, and axis choices can change the message while leaving the code perfectly runnable.
The scope of this experiment
The retained artifact is a sample-style script. It does not provide a verified chart output or a measured account of successful local execution.
It nevertheless captures an interesting application pattern: let a language model suggest analytical directions while keeping the dataset and the calculation available for review. A public demonstration should begin with a small public or synthetic dataset and retain the generated code alongside the chart.