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

ParseBench, the first document OCR benchmark for AI agents, introduces 5 novel metrics for measuring document parsing accuracy. In this video we do a deep dive into parsing charts using our new metric ChartDataPointMatch. Most parsers we tested either skip charts entirely or dump out raw OCR text, neither of which gives an agent usable data. Instead, what is needed is the actual values with their correct series names and axis labels, so that it can be processed in downstream workflows. We annotate up to 10 spot-check data points per chart, each with a numerical value, associated labels, and a tolerance. Charts with explicit value labels must match exactly; charts where values must be read from axes get a 1% tolerance, since pixel-perfect reading is unrealistic. Read more about ParseBench here: https://www.llamaindex.ai/blog/parsebench?utm_medium=video&utm_source=youtube&utm_campaign=2026-apr-