Microsoft and Renmin College Open-Supply ‘Flint’ to Clear up AI Chart-Era Clashes
Microsoft Analysis, in partnership with Renmin College of China’s IDEAS Lab, has open-sourced an intermediate language for information visualization designed to streamline how synthetic intelligence brokers generate graphical charts.
The brand new specification, named Flint, capabilities as a translation layer that permits synthetic intelligence brokers to generate renderable charts for in style visualization libraries like Vega-Lite, ECharts, and Chart.js utilizing single-sentence pure language instructions. The challenge has been made publicly out there on GitHub.
Current visualization libraries provide extremely complicated configuration frameworks. When synthetic intelligence builders use streamlined presets, the ensuing charts usually fail to match the underlying information appropriately. Conversely, when builders job giant language fashions with outputting full, uncooked configurations, the extreme parameters steadily set off architectural conflicts, omissions, or logical inconsistencies.
To bypass these hurdles, Flint decouples the chart’s structural intent from the precise implementation guidelines of the underlying visualization library, lowering the configuration complexity dealt with by the bogus intelligence.
The language splits its specification into two major domains, information definition and chart definition. The information layer annotates the precise semantic kinds of information fields, capturing contextual components like dates, forex, percentages, earnings, geographic international locations, or rankings. The chart layer designates the structure sort and maps these variables to particular visible components just like the horizontal and vertical axes, coloration codes, sizing markers, or sides.
Flint’s built-in compiler evaluates these semantic inputs to routinely decide time-parsing guidelines, aggregation behaviors, numerical formatting, coloration palettes, and total layouts. For example, a monetary revenue variable routinely triggers a diverging coloration scheme to focus on constructive versus unfavourable shifts, whereas a calendar discipline seamlessly inherits the right time-parsing and coordinate axis constructions.
Alongside the core language, the builders launched a devoted flint-chart-mcp server. The server permits synthetic intelligence brokers that assist the Mannequin Context Protocol to assemble and preview information visualizations instantly inside an energetic dialogue, eliminating the necessity to depart the fast chat interface.
In benchmark evaluations, researchers quoted in native Chinese language media in contrast Flint in opposition to DirectVL, a standard method the place synthetic intelligence instantly outputs a whole Vega-Lite configuration. Evaluated by giant language fashions, Flint scored larger complete marks throughout check batches utilizing superior fashions together with GPT-5.1, GPT-5-mini, and GPT-4.1.
