# Chart Understanding, Rebuilt

> Chart understanding now is more accurate, invents less, and is faster, at the same price.

- Canonical: https://www.datalab.to/blog/chart-understanding-rebuilt
- Published: 2026-10-02
- Authors: Leo Chen

**Earlier this year we added a feature that unlocked important information trapped inside charts and diagrams, abundant in business reports and slide decks.**

**Today we're shipping a rebuilt version.**

## Better, and faster

We compared the new version with the previous on an internal evaluation set with thousands of pages. The new version improved on:

- **Accuracy**: outputs score 28% higher on accuracy and 18% higher on completeness than the previous version.
- **Faithfulness**: 56% fewer generated figures contain values unsupported by the source chart, and 93% fewer generated outputs contain incorrect values.
- **Text preservation**: now page text outside a chart never changes. The previous version altered surrounding text on roughly 4% of pages.

It's also faster. We timed both versions on pages with at least one data chart, running [Convert](https://www.datalab.to/platform/processors/convert) in `balanced` mode:

<table class="latency">
  <thead>
    <tr>
      <th rowspan="2">Per page with a chart</th>
      <th colspan="2" class="before-end">Before</th>
      <th colspan="2" class="now">Now</th>
    </tr>
    <tr>
      <th>p50</th>
      <th class="before-end">p90</th>
      <th class="now">p50</th>
      <th class="now">p90</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>Chart understanding step</td>
      <td>34.3 s</td>
      <td class="before-end">47.2 s</td>
      <td class="now"><strong>9.1 s</strong></td>
      <td class="now"><strong>15.3 s</strong></td>
    </tr>
    <tr>
      <td>Full Convert (<code>balanced</code> + chart understanding)</td>
      <td>47.2 s</td>
      <td class="before-end">71.2 s</td>
      <td class="now"><strong>22.6 s</strong></td>
      <td class="now"><strong>45.0 s</strong></td>
    </tr>
  </tbody>
</table>
<p class="latency-note">Lower is better. p50 is the median time it takes to process a page; p90 is the 90th-percentile time.</p>
<!-- Measured on ParseBench chart pages, one page per request, Convert balanced through hosted Chandra API, both chart-understanding versions run on the same Chandra output.-->

## Examples

### U.S. Treasury; Monthly Treasury Statement

Every month the U.S. Treasury publishes an official federal report that provides a precise accounting of tax revenues, expenditures, and national budget. Two pie charts and a stacked bar chart on a monthly are represented as HTML or markdown data tables.

### United Nations; E-Government Survey

The UN publishes global report card on how well governments use digital technology to serve people. The six horizontal bar charts and the marks on the graph are represented as a markdown or HTML data table.

### International Energy Agency; Renewables Report

The IEA publishes Renewables annually as a major assessment of the global renewable-energy market. Combo charts with two axes are represented as a markdown or HTML data table with a text description.

### Japanese Ministry of Health, Labour and Welfare; Regulatory Guidance

This is an official regulatory decision flowchart. It outlines when software qualifies as a medical device under Japan's Pharamaceuticals and Medical Devices Act. Flowchart 4 is represented as a mermaid diagram with an attached description that can be rendered in Markdown.

## Try it out

Chart understanding is an add-on to [Convert](https://documentation.datalab.to/docs/recipes/conversion/conversion-api-overview).

Enable it with `extras=chart_understanding` on the [API](https://documentation.datalab.to/docs/welcome/sdk/conversion#chart-understanding).

The update is live now, and pricing stays the same as before: $3 per 1,000 pages on top of Convert. For on-premise deployments or help evaluating it on your own documents, reach out at support@datalab.to.