Personal Data

The presidio classifier detects personal data, such as names, email addresses, phone numbers, and card numbers, using a Microsoft Presidio analyzer service.

Configuration

classifiers:
  pii0:
    type: presidio
    options:
      endpoint: http://presidio-analyzer:3000
      language: en
      threshold: 0.6
      timeout: PT5S
      labels:
        email: EMAIL_ADDRESS
        person: PERSON

Zilla sends the text to the analyzer's /analyze endpoint. A label is detected when the analyzer finds its entity type at or above the threshold score.

Option	Meaning
endpoint	Base URL of the Presidio analyzer
labels	Your label names, each mapped to a Presidio entity type, such as EMAIL_ADDRESS, PERSON, PHONE_NUMBER, CREDIT_CARD, or US_SSN
threshold	Minimum confidence score, from 0 to 1, for an entity to count. Default 0
language	Language of the text. Default en
timeout	How long to wait for the analyzer. Default PT5S

Running the Analyzer

The analyzer runs as its own service, for example the mcr.microsoft.com/presidio-analyzer container image. If the analyzer can't be reached, times out, or returns an error, the value is treated as detected and rejected.

Rejecting Personal Data in Prompts

north_llm_proxy:
  type: llm
  kind: proxy
  options:
    guardrails:
      gpt-4o-mini:
        request:
          - section: user-text
            value:
              model: classify
              reject:
                pii0:
                  - email

A prompt for gpt-4o-mini whose user text contains an email address is rejected with 400 before it reaches the provider. See LLM Guardrails.

Next Steps

• Classify: how the classify model uses classifiers.
• Semantic: detect content by meaning.
• presidio: the full presidio classifier reference.