The web page and containers#

flask serves a paste-and-go page: paste a prompt, copy the redacted version, paste the model’s reply back, read the restored answer. It needs the web tier (Flask).

pip install "scikit-plots[cleanprompt-web]"
python -m scikitplot.cleanprompt flask --open

The page shows what is not being detected before you paste anything — the one moment that warning is worth reading.

How it keeps your data#

The vault stays on the server, in a bounded, expiring store keyed by an opaque token. The cookie carries the token and nothing else.
Every form carries a CSRF token, compared in constant time.
The session key is configuration. Without CLEANPROMPT_SECRET_KEY a local run uses a per-process key (sessions end with the process), and container mode refuses to start.
Entity detection uses the same engine selection as every other surface: flask --ner --ner-engine nltk --lang en runs exactly that, and a request the machine cannot meet refuses to start with the remedy, instead of serving a page whose banner claims names are being found.

Who can reach it#

Command	Bind and exposure
`flask`	`127.0.0.1`: only this machine
`flask --host 192.0.2.10`	refused: reachable from other machines, no authentication
`flask --host 192.0.2.10 --allow-remote`	allowed, and the banner names the exposure
`flask --docker`	`0.0.0.0` inside a container, which a published port needs

Debug mode is loopback only. Flask’s debug mode serves an interactive debugger that runs Python typed into the browser. --debug is therefore refused with --docker or any non-loopback --host, whatever else is acknowledged: --allow-remote accepts that the page is reachable, not that code execution is.

This is Flask’s development server: right for one person on one machine, wrong for a shared deployment. A shared deployment needs authentication, TLS and a production server in front of it, which this command does not provide.

Container files#

python -m scikitplot.cleanprompt docker --write ./deploy
python -m scikitplot.cleanprompt docker --write ./deploy --with-ner

This writes a Dockerfile, a docker-compose.yml, a .dockerignore and a README.md. Everything in them is derived from the runtime:

Loopback publishing. Both the compose file and the docker run line publish 127.0.0.1:PORT:PORT. Docker’s -p PORT:PORT without an address publishes on every interface of the host — do not shorten it.
The model it installs is the model it runs. With --with-ner the image downloads the model the runtime resolves (en_core_web_sm by default) and starts with --ner --ner-engine spacy --ner-model naming that same model, so the two cannot drift apart.
A secret key with no default. CLEANPROMPT_SECRET_KEY is required; a default would be a published secret shared by every deployment.
Least privilege. The image runs as a non-root user; the compose file mounts the filesystem read-only, drops every capability and sets no-new-privileges. .dockerignore keeps vault files out of the image.

export CLEANPROMPT_SECRET_KEY="$(python -c 'import secrets;print(secrets.token_hex(32))')"
docker compose -f deploy/docker-compose.yml up --build
# then open http://127.0.0.1:5000

From Python, container_files returns the same files as strings and takes language and model_size for the model layer.

The terminal session#

cli is the same loop in a terminal, holding one vault for the session:

$ python -m scikitplot.cleanprompt cli
CleanPrompt interactive session
  Paste your text, then press Ctrl-D on a blank line to submit.
  Commands: :hide TERM   :allow TERM   :suggest   :why   :again   :quit