PII Detection

The PII detection endpoint identifies personally identifiable information (PII) in text and returns structured spans with a ready-to-use redacted string. It is built on OpenAI Privacy Filter — a 1.5B sparse-MoE token classifier.

Detected categories: account_number, private_address, private_email, private_person, private_phone, private_url, private_date, secret

Quick start

curl https://api.casola.ai/openai/v1/pii \
  -H "Authorization: Bearer $CASOLA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/privacy-filter",
    "input": "My name is Harry Potter and my email is harry.potter@hogwarts.edu"
  }'

Response:

{
  "object": "pii.detection",
  "model": "openai/privacy-filter",
  "results": [
    {
      "spans": [
        { "label": "private_person", "start": 11, "end": 23, "score": 0.9987, "text": "Harry Potter" },
        { "label": "private_email",  "start": 40, "end": 65, "score": 0.9994, "text": "harry.potter@hogwarts.edu" }
      ],
      "redacted_text": "My name is [PRIVATE_PERSON] and my email is [PRIVATE_EMAIL]"
    }
  ]
}

Batch inputs

Pass an array of strings to process multiple texts in a single request:

{
  "model": "openai/privacy-filter",
  "input": [
    "Call me at 555-867-5309.",
    "Ship to 1 Infinite Loop, Cupertino, CA 95014."
  ]
}

The results array matches the order of the input array.

Response fields

Field	Type	Description
results[].spans	array	Detected PII spans (see below)
results[].redacted_text	string	Input text with each span replaced by [LABEL]

Each span has:

Field	Type	Description
label	string	PII category (e.g. private_email)
start	integer	Character start offset (inclusive)
end	integer	Character end offset (exclusive)
score	float	Average token-level confidence for this span
text	string	The matched substring

Async mode

For large inputs or batch workloads, add "async": true to get a job ID you can poll:

curl https://api.casola.ai/openai/v1/pii \
  -H "Authorization: Bearer $CASOLA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"openai/privacy-filter","input":"...","async":true}'

Poll the returned job ID at GET /api/jobs/{id}.

Model notes

• Context limit: the underlying banded-attention model supports up to 512 tokens per call; longer inputs are silently truncated. Split very long documents into paragraphs before processing.
• Language support: optimized for English. Cross-lingual detection is partially supported for proper names and email/URL/phone patterns.
• Not a generative model: Privacy Filter runs as a pure token classifier — it does not generate text and has no system prompt.