Operator console and HTTP API

Install and start the local application:

python -m pip install -e ".[web]"
indad-web --datasets ./datasets --host 127.0.0.1 --port 8000

Open http://127.0.0.1:8000. No Node installation, build step, external fonts, or frontend service is required. HTML, CSS and JavaScript ship inside the Python package. python -m indad.server is equivalent to the console entry point.

The console is a local workspace with no accounts or authentication. The default bind address is loopback. Files are scoped to the storage folder configured at startup. Training and prediction use the server's compute device.

Operator workflow

1. Create or select a collection from the sidebar.
2. Import images as healthy training, healthy inspection, defective inspection, or defect masks. Choose a defect type when relevant. Imports preserve original filenames and bytes; existing files are never overwritten. Each batch is limited to 500 files and 100 MB.
3. Browse samples, search filenames, filter by collection, and review the selected image and its mask. Arrow buttons move between samples. The gallery loads 24 samples at a time and uses thumbnails.
4. Use Correct label or collection or Exclude sample in the sample panel, record the reason, and save. Review or undo edits in Change history. Excluded images and unused masks are preserved in the dataset's .indad folder.
5. Open Quality checks to see operation-specific readiness and actionable findings. Clicking a finding's sample path opens the relevant image.
6. Open Inspection baseline, choose a core model, and train. The task runs in the background; you can continue reviewing samples. Select an inspection image to see the model input, anomaly map, overlay, and numerical score.

Refresh rescans changes made on disk or through another client. Export manifest saves the same deterministic JSON inventory used by indad-data inspect. Annotation coverage counts linked mask files; mask validity is checked separately.

One baseline is held per server. It survives page reloads but is replaced when a new baseline starts and is lost on restart. The interface matches dataset contents and model settings to the run; the API rejects predictions if dataset contents have changed. Concurrent training/prediction requests receive a conflict response. A failed training run includes its error in the run status.

Agent workflow

The API documentation is at /docs; its machine-readable schema is at /openapi.json. Dataset operations call the same Python functions as the CLI.

Method	Endpoint	Purpose
GET	/api/health	Server status and version
GET	/api/datasets	List local dataset names
POST	/api/datasets	Create a dataset with {"name":"parts"}
GET	/api/datasets/{name}/report	Inventory, checksums, findings and readiness
POST	/api/datasets/{name}/import	Multipart files, split, label
GET	/api/datasets/{name}/image?path=...&thumbnail=true	PNG preview of a dataset image or mask
GET	/api/datasets/{name}/changes	Edit and undo history, newest first
POST	/api/datasets/{name}/samples/revise	Relabel, move, or exclude a sample
POST	/api/datasets/{name}/changes/{id}/undo	Reverse a recorded edit
POST	/api/runs	Start background training
GET	/api/runs/current	Current run, or null before training
GET	/api/runs/{id}	Run status and dataset fingerprint
POST	/api/runs/{id}/predict	Inspect a test image with {"path":"test/good/001.png"}

For example:

curl -X POST http://127.0.0.1:8000/api/datasets \
  -H 'Content-Type: application/json' -d '{"name":"parts"}'

curl -X POST http://127.0.0.1:8000/api/datasets/parts/import \
  -F 'split=train' -F 'label=good' \
  -F 'files=@captures/healthy_001.png' -F 'files=@captures/healthy_002.png'

curl http://127.0.0.1:8000/api/datasets/parts/report

# Add inspection images before starting a baseline.
curl -X POST http://127.0.0.1:8000/api/runs \
  -H 'Content-Type: application/json' \
  -d '{"dataset":"parts","method":"patchcore","device":"cpu"}'

Training returns HTTP 202 with a run ID. Poll the run until status is ready or failed. Other states are training, predicting, and stale. Run options are patchcore, padim, or spade, with cpu or cuda. The standard baseline uses ResNet18, a 224 × 224 crop, and seed 0. The first run may download pretrained weights.

Prediction returns the sample path, run ID, image anomaly score, score-map minimum and maximum, and PNG data URLs for original (the model input), heatmap and overlay. Heatmap normalization is per image. Scores are not pass/fail decisions.

File and validation errors return a non-2xx response with a JSON detail. Unready or stale baselines and busy inspection tasks return HTTP 409; oversized imports return 413. Invalid typed request fields return 422. Inspect the report's readiness flags before requesting a run.

Sample curation requests

POST /api/datasets/{name}/samples/revise accepts:

{
  "path": "test/scratch/001.png",
  "fingerprint": "fingerprint from the latest report",
  "reason": "Reviewed defect type",
  "split": "test",
  "label": "dent",
  "exclude": false
}

Set exclude to true to archive the sample and its mask. For exclusion, split and label are ignored. Every edit requires a nonempty reason of up to 500 characters. The result includes a change ID, timestamp, reason and before/after file paths with checksums.

Undo takes {"fingerprint":"latest report fingerprint"}. Restore is refused if archived files changed or their original paths now contain other captures. Undo successive edits to the same image in reverse order. Stale fingerprints return HTTP 409. No operation overwrites existing images or masks.

The history endpoint marks previously undone edits with undone: true. The archive/history directory must accompany dataset backups to preserve undo.

Development

python -m pip install -e ".[web,dev]"
make lint
python -m pytest -m 'not export'

API tests use deterministic model backbones without downloading weights. Optional browser tests require playwright and python -m playwright install chromium. Static files live in indad/web/; dataset operations remain in indad/workspace.py.