privacy-filter

FastAPI and Ray Serve app for openai/privacy-filter.

The service loads the Hugging Face token-classification model in a Ray Serve actor, finds PII spans, and can return either spans only or masked text.

Setup

make setup

This creates .venv/, installs dependencies with uv, and installs the local pre-commit hooks.

Run Locally

make serve

The service listens on 0.0.0.0:8080 by default. Override with HOST or PORT:

PORT=8000 make serve

Ray Serve import path for charts or serveConfig:

privacy_filter.service:deployment

Docker

make docker-build
make docker-run

Equivalent commands:

docker build -t privacy-filter:local .
docker run --rm -p 8080:8080 privacy-filter:local

Configuration

MODEL_ID changes the model loaded by the Serve actor. It defaults to openai/privacy-filter. MASTER_KEY sets the API key required by POST /privacy-filter.

MASTER_KEY=sk-default MODEL_ID=openai/privacy-filter make serve

API

Health checks:

curl http://localhost:8080/healthz
curl http://localhost:8080/readyz

Filter text:

curl -X POST http://localhost:8080/privacy-filter \
  -H "content-type: application/json" \
  -H "x-pf-api-key: $MASTER_KEY" \
  -d '{
    "items": ["Email jane@example.com or call 555-0100."],
    "mask": true,
    "threshold": 0.5
  }'

POST /privacy-filter requires the x-pf-api-key header. Its value must match the service MASTER_KEY environment variable. Health check endpoints do not require an API key.

Response shape:

{
  "results": [
    {
      "masked_text": "Email [EMAIL] or call [PHONE].",
      "spans": [
        {
          "category": "email",
          "text": "jane@example.com",
          "start": 6,
          "end": 22,
          "score": 0.99
        }
      ]
    }
  ],
  "model_id": "openai/privacy-filter",
  "num_items": 1
}

Swagger UI is available at http://localhost:8080/docs when the service is running.

Development

make test
make lint