AI
Microsoft's Flint: An AI-Friendly Visualization Language That Makes Better Charts Simple
A new open-source language lets AI agents generate polished, expressive charts from compact specifications that are both human-readable and machine-friendly.
Microsoft Research Blog
Creating publication-quality charts typically demands wrestling with dozens of low-level design choices—axis scales, date formats, label sizing, color palettes. Traditional visualization libraries offer control over all these details, but the cost is bloated, fragile specifications that are hard for humans to review or for AI systems to generate reliably.
Flint, a new intermediate language from Microsoft Research, bridges that gap by letting AI agents produce attractive, semantically sound charts from minimal, editable specifications. The compiler then handles the messy details automatically.
How Flint Simplifies Chart Creation#
Rather than forcing users or agents to specify low-level parameters, Flint accepts a compact JSON description of the data and the desired visualization. The system infers semantic meaning—recognizing whether a column holds dates, currency, percentages, or geographic names—and uses that context to derive parsing rules, appropriate scales, formatting conventions, color schemes, and layout automatically.
This approach frees AI systems from having to reason about brittle, library-specific configuration. A human can read and edit a Flint spec in seconds, while an LLM can generate one far more reliably than it can produce a full Vega-Lite or ECharts specification.
Multi-Backend Compilation#
One of Flint's key strengths is backend independence. The same compact specification can compile to Vega-Lite, Apache ECharts, or Chart.js without rewriting the chart. This lets teams swap rendering engines based on which best fits their needs while keeping the chart intent unified.
Proven Performance with LLMs#
Microsoft tested Flint against a baseline approach (DirectVL) that asks language models to generate complete Vega-Lite specs directly. Across GPT-5.1, GPT-5-mini, and GPT-4.1, Flint consistently outperformed the baseline in LLM-judge evaluation, with scores ranging from 16.27 vs. 15.91 (GPT-5.1) to 15.91 vs. 15.34 (GPT-4.1). The system proved so robust that it now powers Data Formulator, a separate Microsoft Research tool for AI-assisted analytics and visualization.
