Understand Chart (ChartGemma)

@ahmed_masry1 VersionGenerate

Takes a chart Image and a question String, runs a PaliGemma-based vision-language pass over the pixels, and emits a single answer String reading values, comparing series, or summarising the trend.

How it fits

Typical backends

Ask a question about an uploaded chart and publish the answer as JSON. The file source supplies the chart figure and the question is wired in alongside.

Read charts off a live HTTP image feed and drive the question from an expression node so the prompt can be templated per deployment.

Caveats

I/O contractThe chart Image and the question String arrive together and the question conditions the answer; an empty question is replaced internally with a generic chart-description prompt so the component always answers.
FallbackAn empty image yields the empty String without running the model, and any inference error is caught and also returned as an empty answer rather than crashing.
AccuracyThe model is trained on rendered chart figures; on ordinary photographs or heavily distorted chart captures answer quality drops, and the recommended beam width trades latency for accuracy.
LatencyThis is a multi-billion-parameter vision-language model run with beam search; CPU inference is impractical and a GPU with sufficient memory is required for live use.
State lifetimeThe model and processor are loaded once at startup and reused for every call; decoding settings are fixed at that point and a restart is needed to change them.

Versions

ee74025blatestdefault
Automated release
6/28/2026