| name | privacy-mode |
| description | Install and configure privacy mode for local PII anonymization. Activated by natural language ("enable privacy mode", "anonymize everything", "don't send data externally") or via /privacy-mode. Downloads the OpenAI Privacy Filter model and sets up the maskzone/ pipeline. |
Privacy Mode
No PII leaves the machine.
You have been invoked to set up, toggle, or verify privacy mode.
Follow these steps exactly.
Step 0 โ Check current state
Run:
python tools/privacy_filter.py --status
This tells you:
- Whether privacy mode is currently active or inactive
- Whether dependencies are installed (
.privacy-deps-ok exists)
- Whether
maskzone/ exists and how many files are pending
Based on the result, follow the appropriate flow below:
| Current state | User wants | โ Flow |
|---|
| inactive + deps NOT installed | enable | First-time activation |
| inactive + deps installed | enable | Reactivation |
| active | disable | Deactivation |
| active | enable | Already active โ just confirm to the user |
Flow A โ First-time activation
A1 โ Explain and confirm
Tell the user what will happen:
- Privacy mode anonymizes all PII locally before any data reaches external agents
- Documents go in
maskzone/ instead of raw/
- A local model (OpenAI Privacy Filter, Apache 2.0) runs on-device via ONNX Runtime
- Originals stay in
maskzone/ after processing โ the user removes them manually when ready
- First-time setup requires downloading the model (~17 GB, cached in
~/.cache/huggingface/)
and installing Python dependencies
Ask the user for confirmation before proceeding.
A2 โ Install dependencies
Run:
pip install -r requirements-privacy.txt
If pip is not found, try pip3 or python -m pip.
If the installation fails:
- Report the exact error to the user
- On Windows with permission errors, suggest:
pip install --user -r requirements-privacy.txt
- Do NOT proceed to A3 if installation failed
A3 โ Verify imports
Run:
python -c "import transformers, docx, pdfplumber, pptx, openpyxl, bs4; print('ALL_PRIVACY_DEPS_OK')"
If the output contains ALL_PRIVACY_DEPS_OK, proceed to A4.
If any import fails, report which package failed and re-run
pip install <package> for that specific package. Then verify again.
Note: onnxruntime and optimum are optional (ONNX acceleration). The script
automatically falls back to standard PyTorch inference if they are unavailable.
A4 โ Download model and activate
Run:
python tools/privacy_filter.py --setup
This downloads the model, runs a test inference, creates .privacy-deps-ok,
sets can_anonymize_pii to true in wiki/.protect, and creates maskzone/.
If setup fails:
- Network error โ check internet connection, retry
- Model error โ report and suggest retrying
- Do NOT proceed if setup failed
A5 โ Report
Tell the user:
- Privacy mode is now active
- Place files in
maskzone/ instead of raw/
- All files will be anonymized locally before any external agent or service can access them
- Originals stay in
maskzone/ โ remove them manually when no longer needed
- To disable: say "disable privacy mode"
- To check status:
python tools/privacy_filter.py --status
Flow B โ Reactivation
Dependencies and model are already installed. No download needed.
B1 โ Explain and confirm
Tell the user:
- Privacy mode will be reactivated
- Dependencies and model are already cached โ no reinstallation needed
- Documents will go in
maskzone/ and be anonymized locally
Ask the user for confirmation before proceeding.
B2 โ Toggle flag
Run:
python tools/wiki_protect.py --toggle can_anonymize_pii
B3 โ Report
Tell the user privacy mode is active again. Same usage as A5.
Flow C โ Deactivation
C1 โ Explain and confirm
Tell the user:
- Privacy mode will be disabled
- Ingest workflow returns to normal (files go in
raw/)
- Dependencies and model stay cached (no re-download needed to reactivate)
Ask the user for confirmation before proceeding.
C2 โ Toggle flag
Run:
python tools/wiki_protect.py --toggle can_anonymize_pii
C3 โ Report
Tell the user privacy mode is disabled and can be reactivated at any time.
