deployment-and-contrib

deployment-and-contrib

Use this sub-skill when a docTR task is about an optional surface around the core OCR package rather than a plain DocumentFile -> predictor -> result workflow:

• exposing OCR/KIE through a FastAPI service;
• adapting the Streamlit demo flow;
• choosing Docker images or Docker GPU settings;
• installing or debugging optional html, viz, or contrib extras;
• using doctr.contrib.ArtefactDetector;
• loading from or sharing to the Hugging Face Hub.

Start by deciding whether the user actually needs a service/demo. For scripts, notebooks, batch jobs, or one-off research experiments, route them to the package API sub-skills instead of starting a long-lived service.

Read order

1. For API, demo, and Docker deployment shape, read references/api-and-demo.md.
2. For optional extras, ArtefactDetector, and Hugging Face Hub loading/pushing, read references/contrib-and-hub.md.
3. For missing extras, MIME/route errors, GPU/Docker issues, and Hub authentication failures, read references/troubleshooting.md.

Route away

• Use the core OCR/KIE sub-skill for normal ocr_predictor/kie_predictor Python API work, predictor parameters, batching, and result interpretation.
• Use the document IO/export sub-skill for DocumentFile, HTML input, visualization/export methods, and output object semantics.
• Use the models/customization sub-skill for architecture selection, model factories, ONNX export, precision/compile/device optimization, and custom weights outside Hub packaging.
• Use the CLI/scripts sub-skill for installed doctr-cli and bundled batch helpers.
• Use the datasets/training sub-skill before Hub publishing when the model still needs training, evaluation, or artifact validation.

Safety boundaries

• Do not start a web server, Streamlit app, Docker container, browser session, interactive Hub login, or Hub push unless the user explicitly asks for that side effect.
• Do not ask for, echo, save, or invent credentials. If a Hub operation needs authentication, stop and ask the user to provide an already-configured token/session or to approve an interactive login.
• Do not claim full service, Docker, GPU, or Hub verification from static inspection alone. Treat the recipes here as deployment guidance until the requested runtime check is actually run.
• Keep service ports, file-size limits, authentication, concurrency, and model cache behavior explicit when moving from a demo/template to production.