Ecore Inspection Target Planning

Ｑuick start

git clone https://github.com/ShellWee/inspection-demo.git
cd inspection-demo

Open notebooks/Launch_Demo.ipynb in VS Code or Jupyter and execute its cells in order. The notebook kernel needs an approved Python 3.10+ interpreter; setup creates a separate locked Python 3.12 application environment. The built React UI is included: Node.js is not required to run. If you need a notebook editor, install JupyterLab in your approved editor/kernel environment, not in the app environment that uv sync manages.

The notebook installs dependencies, verifies assets, starts the local server, checks status, optionally submits a query and stops the server. Setup and start do not call an OpenAI model.

• Assets: approximately 943 MB, downloaded separately from the private HF dataset revision pinned in assets.lock.json. Obtain read access and use hf auth login, or point ASSETS_DIR to an existing complete asset bundle.
• Local Jupyter/VS Code: open the printed localhost URL for the full three-stage UI.
• Colab: use the notebook's loopback HTTP query cells. This is not public web hosting, and no tunnel or background keep-alive workaround is created.
• CPU is the default; CUDA is optional on compatible Linux hosts. Linux currently uses the locked CUDA-enabled Torch distribution even in CPU mode, so dependency downloads are substantially larger than the asset bundle. Disk requirements also include the extracted model and copied graph index.
• The app still uses the existing single-user ownership model. Do not expose its port publicly. This release does not implement the proposed remote ZeroGPU service or change cloud billing.

See step-by-step notebook instructions and source/asset notices.

Updating an existing checkout

git pull --ff-only

Stop the existing demo first, preserve your uncommitted changes, then rerun setup and asset verification from the notebook. A checkout is identified by its Git commit; RELEASE_MANIFEST.json checks the shipped bytes. Do not copy a Colab/Linux virtual environment onto Windows.

Maintainer checks (Python 3.12+, no model/API calls)

python -m unittest discover -s scripts/tests -v
python scripts/verify_release.py
python scripts/repository_manifest.py --check

After editing source or the UI, rebuild the frontend with the locked package manager environment, review the diff, clear notebook outputs, then regenerate the inventory with python scripts/repository_manifest.py --write. The historical SOURCE_MANIFEST.json is preserved as research provenance. Include the current inventory with each release; do not commit downloaded assets or state directories.

Space configuration

• Hardware: l4x1
• Visibility: private
• Read-only volume: hf://datasets/<namespace>/inspection-demo-ecore-assets-v1:/mnt/default-assets:ro
• One Uvicorn worker on port 7860

The entrypoint also supports HF_ASSET_DATASET plus an HF_TOKEN secret as a fail-closed fallback when native Dataset mounting is unavailable. The normal deployment uses the read-only volume and does not expose a Hub token to the container.

After hf auth login, deploy both private repositories and attach the L4 with:

.\scripts\deploy_hf.ps1 -Namespace <hugging-face-username>

An l4x1 Space requires Hugging Face prepaid credits. The deployment script reserves the Space before uploading the large asset bundle and stops immediately on any non-zero CLI exit code.

No OpenAI API key is persisted. Run state, events, trajectories, and replays use ephemeral storage and disappear when the Space restarts.

Local verification

The UI now uses a simplified three-stage inspection workspace. See refinement notes for changes, defect fixes and cost-limited test instructions. Internal pipeline identifiers remain in engineering artifacts, not the main user interface.

uv sync --frozen --extra dev
uv run pytest
pnpm --dir frontend install --frozen-lockfile
pnpm --dir frontend test
pnpm --dir frontend build
docker build -t inspection-demo-hf .