Workshop: Multi-Series Line Chart with a Legend

Workshop two of five. Steps up from the bar chart to three overlapping series — where vision LLMs swap series and produce confidently wrong output. Calibrated handles it the same as a single line: one series at a time.

The practice chart

Open this chart in DataFromChart →

Three product lines (A blue, B green, C orange) monthly over a year. 36 points. Y-axis “units sold (thousands)” 0-130; x-axis Jan-Dec.

Target: three series, twelve (month, value) pairs each.

The trick: extract one series at a time

Finish one series completely before starting the next. All twelve Product A points first, then Product B. Don’t interleave.

Why: with all series in one extraction, it’s easy to lose track of which point belongs to which — especially at crossings. The data layer can’t separate them after the fact.

DataFromChart supports per-series extraction natively: three points groups, series name as label.

Step 1: open the chart and create the first series

Open the chart, advance to POINTS, create a group “Product A”. The app defaults to blue — usually matches.

Step 2: place Product A points

Click each of the twelve blue-line points left to right. Zoom in — land on the marker.

Twelve clicks. Blue points should sit on top of the blue data.

Step 3: calibrate the axes

Y-axis: 0 at the bottom gridline, 120 near the top. X-axis: Jan (month 1) and Dec (month 12), or skip x-calibration if you’ll re-label by month name in the export.

Step 4: repeat for Products B and C

Create “Product B”. Twelve green-line points. Then “Product C”, twelve orange. Don’t re-calibrate — axes apply to all series.

Total: 36 clicks, two calibrations, three named groups.

Step 5: export

XLSX or CSV. All 36 points with series names in long format:

series,month,value
Product A, Jan, 31.2
Product A, Feb, 36.0
...
Product B, Jan, 42.4
...

Long format is what Plotly, ggplot, and Vega-Lite prefer.

Answer key

Ground-truth values:

Month	Product A	Product B	Product C
Jan	39.6	49.8	44.5
Feb	37.4	48.8	42.0
Mar	39.2	52.8	42.9
Apr	41.6	55.7	39.3
May	41.3	54.2	42.0
Jun	42.8	55.9	45.6
Jul	39.4	57.2	47.3
Aug	36.0	61.9	52.1
Sep	34.1	65.2	51.2
Oct	36.9	64.1	54.2
Nov	37.2	69.9	56.1
Dec	36.3	67.1	57.9

Per-series MAE. Aim for under 1.5. Above 3 = calibration drift; above 5 with no calibration issue = a point on the wrong series, usually at a crossing.

Common mistakes

• Swapping series at crossings. When blue and green cross around May, it’s easy to click the green line while extracting Product A. Named groups make this visible — a Product A point on the green line is an obvious mismatch. Re-check at crossings.
• Forgetting to switch active groups. Clicking while “Product A” is still active adds to Product A. Check the active group label before each set.
• Inconsistent x-positions across series. All three should share x-positions per month, but eyeballing “Jan” differently each time misaligns your Jan values. Set x-snap to month positions if precision matters.

How this compares to AI

Multi-series is where AI fails most visibly. From our benchmark:

• Claude Sonnet 4.6: 18% MAE, swaps in 5 of 12 months.
• GPT-4o: 22% MAE, consistent swap on the B/C crossing.
• Gemini 2.5 Pro: 20% MAE.

All three could name the series from the legend. None could reliably tie a name to the right line at each x. Per why AI gets chart data wrong: tracking lines through crossings needs sustained spatial reasoning vision encoders aren’t built for.

Calibrated extraction beats all three by 15-20 percentage points in ~8 minutes.

Next

• Workshop 3: Dense Scatter Plot with Auto-Detect — 250 points in 90 seconds with color-based auto-extraction.
• Workshop 4: Log-Scale Chart Without Arithmetic Mistakes — log axes, the failure mode AI never solves.
• All five workshops + practice datasets — full hub.