Atelier turns a published paper into a FindSim experiment definition. Given a PDF and a figure or panel to target, the pipeline reads the paper, builds the experiment description, digitizes the plotted data, and writes a FindSim record.

pipeline.py is the single orchestrator that runs all of them.

Python 3.11 or newer (developed on 3.14).

git clone https://github.com/<your-org>/FSW-AI.git
cd FSW-AI
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Every stage uses one LLM backend, set with --backend:

Use --model to override the model for the chosen backend.

The pipeline needs two things at minimum:

1. A PDF of the paper, passed with --pdf, or a directory of PDFs passed with --pdf-dir.
2. A target, meaning which figure or panel to build an experiment from. Pass it inline with --target, or list many targets in a manifest file.

Two optional inputs enable the model builder stage. Supply both or neither:

3. --sim-model, a simulation model file, either kkit .g or HillTau .json.
4. --sim-map, the matching model map JSON, for example NOS_ma_map.json.

cd FSW-AI/pipeline
python pipeline.py \
    --pdf /path/to/Abrams1991.pdf \
    --target "Figure 2:A" \
    --output-dir out/ \
    --backend cc

--target is repeatable. It accepts "Figure 2" for a whole figure or "Figure 2:A" for one panel. Omit it entirely to process every figure the paper contains.

A manifest is a JSON list of targets. PDFs are resolved as <pdf-dir>/<paper>.pdf.

[
  {"paper": "Abrams1991",   "figure": "Figure 2", "panel": null},
  {"paper": "Alano2010",    "figure": "Figure 3", "panel": "B"},
  {"paper": "Fujiwara2014", "figure": "Figure 3", "panel": "C"}
]

python pipeline.py \
    --manifest manifests/list1.json \
    --pdf-dir /path/to/pdfs \
    --output-dir out/ \
    --backend api \
    --parallel 4

Example manifests are in FSW-AI/pipeline/manifests/. To build one from a directory of ground truth filenames such as Abrams1991_Fig2.json:

python manifest_from_groundtruth.py --groundtruth-dir /path/to/Expts --output targets.json

Size --parallel to your API rate limit, not your CPU count. Every worker draws from the same account wide budget.

One directory per paper, one subdirectory per figure or panel:

out/
└── Abrams1991/
    ├── paper.json                  # stage 1
    ├── figures/                    # extracted figure images
    └── figure2A/
        ├── experiment.json         # stage 2
        ├── data.json               # stage 3
        ├── extracted_data.csv      # digitized points
        ├── extracted_plot.png      # digitizer overlay
        ├── comparison_plot.png     # digitized vs original
        ├── experimentMap.json      # stage 4, only with --sim-model/--sim-map
        └── findsim_preview.json    # final FindSim record

Run python pipeline.py --help for the full list.

If a FindSim checkout sits beside this repository, its FindSim-Schema.json is found automatically. Otherwise point at it directly:

export FINDSIM_SCHEMA=/path/to/FindSim/FindSim-Schema.json

Without it, folding falls back to the Pydantic mirror in core/FindSim_schema.py.

cd FSW-AI/pipeline
pytest -m unit

Unit tests mock every LLM call. Tests marked integration need a live backend, and those marked docling run real PDF layout extraction. Tests that compare against FindSim itself skip unless FINDSIM_DIR points at a checkout.

separate_panels_by_whitespace.py    # standalone panel splitter, imported by the pipeline
WIP docs/                           # flowchart and curation notes
FSW-AI/pipeline/
├── pipeline.py                     # main orchestrator
├── experimentBuilder.py            # stage 2 CLI
├── dataBuilder.py                  # stage 3 CLI
├── modelBuilder.py                 # stage 4 CLI
├── fold.py                         # stage 5, merges data into the FindSim record
├── core/                           # library modules for every stage
├── agents/                         # LLM prompts, one CLAUDE.md per agent
├── manifests/                      # example target lists
└── templates/                      # pdf_annotator UI template

The algorithms here are built by the authors, and the code is generated by the authors with the help of LLMs, specifically Claude's Sonnet 4.5 and Sonnet 5.0.