Vision medium
Chart Data Extraction — Quarterly Revenue Bar Chart
Vision-capable models extract structured data from a bar chart image. Scored on numerical accuracy and schema compliance.
Published June 8, 2025
Goal
Evaluate vision models' ability to extract quantitative data from charts for downstream automation.
Models Compared
- GPT-4o OpenAI 2024-11-20
- Claude 3.5 Sonnet Anthropic 20241022
- Gemini 2.0 Flash Google gemini-2.0-flash-001
Exact Prompts
system system
Extract data from the chart image. Return valid JSON matching the provided schema exactly. user user
Extract all data points from this bar chart into JSON:
```json
{
"quarters": ["Q1 2023", ...],
"series": [
{ "name": "string", "values": [number, ...] }
]
}
```
[Chart image attached: revenue-bar-chart-q1q4.png]
Parameters
- Temperature
- 0
- Max Tokens
- 2048
- Top P
- 1
Context
A 1200×800 PNG bar chart showing quarterly revenue for 4 product lines across 2023–2024. Ground truth values provided separately for scoring.
Outputs
GPT-4o
Extracted 16 data points. 15/16 within 2% of ground truth.
Claude 3.5 Sonnet
Extracted 16 data points. 16/16 exact match.
Gemini 2.0 Flash
Extracted 16 data points. 14/16 within 2%. Swapped Q3/Q4 for Product C.
Evaluation
- Method
- Automated comparison against ground truth CSV
- Rubric
- Numerical accuracy (70%), schema compliance (20%), completeness (10%)
- Human Review
- No — automated only
Detailed Analysis
Chart Specifications
- Format: PNG, 1200×800, sRGB
- Chart type: Grouped bar chart, 4 series × 4 quarters
- Value range: $1.2M – $8.7M
- Ground truth: Provided as
revenue-ground-truth.csvin downloads
Verdict
Claude 3.5 Sonnet achieves perfect extraction on this chart type. Gemini 2.0 Flash is fastest and cheapest but less reliable on adjacent bar values.