PlotPick

Live demo: plotpick.streamlit.app

AI-powered extraction of numerical data from scientific figures.

Upload images, PDFs, or ZIP archives. Each figure is sent to Claude's vision API with a structured extraction prompt. Results are displayed as tables and can be exported in multiple formats.

Features

• PDF figure detection -- automatically finds and crops individual figures and tables from multi-page PDFs using caption detection
• Batch processing -- upload multiple files at once
• Structured extraction -- reads boxplots, bar charts, and line plots with biomarker, group, timepoint, and summary statistics
• Export formats -- Markdown, Excel, CSV, JSON

Architecture

flowchart TD
    A["📁 Upload images, PDFs, or ZIPs"] --> B["🔍 Automatic figure detection<br/>from PDFs"]
    A --> C["🖼️ Image gallery"]
    B --> C
    C --> D["☑️ Select figures to extract"]
    D --> E["🔑 Send to Claude Vision API"]
    E --> F{"🤖 AI reads the figure"}
    F -->|✅ Success| G["📊 Structured data table<br/>with confidence score"]
    F -->|❌ Error| H["⚠️ Flag & continue"]
    G --> I["📤 Export"]
    I --> J["📗 Excel"]
    I --> K["📄 CSV"]
    I --> L["🔬 LaTeX"]
    I --> M["📦 JSON / R / Markdown"]

Loading

Quickstart

1. Install dependencies:

   pip install -r requirements.txt
   

2. Add your Anthropic API key to .streamlit/secrets.toml:

   ANTHROPIC_API_KEY = "sk-ant-..."

3. Run the app:

   streamlit run streamlit_app.py
   

Validation dataset

The validation/ folder contains scripts to build a ground-truth dataset for benchmarking PlotPick against structured table data from the same papers.

1. Find candidates -- query PMC for open-access articles that cross-reference a table and a figure presenting the same data:

   python validation/find_validation_papers.py
   

2. Download PDFs + extract tables -- fetch the PDF and parse structured table data from the PMC XML (ground truth):

   python validation/download_ground_truth.py [--limit N]
   

3. Match table-figure pairs -- identify which Table N corresponds to Figure M, extract the figure image, and filter for numeric tables:

   python validation/match_pairs.py [--limit N]
   

4. Run benchmark -- send each figure through Claude's vision API and compare extracted values against ground-truth tables:

   python validation/run_benchmark.py [--model sonnet|haiku] [--limit N]
   python validation/run_benchmark.py --report   # regenerate report only
   

Output:

• validation/candidates.csv -- candidate paper metadata
• validation/pdfs/ -- downloaded PDFs
• validation/tables/ -- structured table data (JSON, one file per paper)
• validation/pairs.json -- matched table-figure pairs (199 pairs)
• validation/figures/ -- extracted figure PNGs
• validation/results/ -- per-pair extraction results
• validation/benchmark_report.md -- aggregate accuracy metrics

Requirements

• Python 3.12+
• An Anthropic API key