Deep Dive into ChartDataPointMatch: A New Metric for Evaluating Parsing Accuracy on Charts

Short summary

LlamaIndex introduces ParseBench, the first OCR benchmark for AI agents, featuring ChartDataPointMatch—a metric measuring whether parsers extract actual values, labels, and axis names rather than raw OCR text. The metric requires 10 spot-checked data points per chart, with exact-match rules for labeled values and 1% tolerance for axis readings. This addresses a critical gap: most parsers either skip charts or output unusable data.

• •ParseBench introduces ChartDataPointMatch metric for evaluating document parsing accuracy in AI agents
• •Metric requires extraction of actual values with series names and axis labels, not raw OCR text
• •Includes tolerance rules: exact matches for labeled values, 1% tolerance for axis readings

Generated with AI, which can make mistakes.