PII Firewall

Stop leaking PII into LLM APIs without breaking conversation context.

PII Firewall intercepts sensitive data before it reaches OpenAI, Anthropic, or any model provider, then restores it transparently in the response.

User text  ──►  [ PII Firewall ]  ──►  Sanitized prompt  ──►  LLM
                      │                                          │
                 Secure vault                             Model response
                      │                                          │
                      └──────────  Re-hydrated reply  ◄─────────┘

Standard redaction breaks context — replacing "John" with [REDACTED] means the model can no longer refer to the person by name. PII Firewall uses a stateful Detect → Anonymize → Rehydrate flow so the LLM sees pseudonyms while the user gets real answers.

---

Try it in 30 seconds (no ML models needed)

pip install pii-firewall

from privacy_firewall import create_firewall

firewall = create_firewall("healthcare", detector_backend="regex")

result = firewall.secure_call(
    text="Patient John Doe, SSN 123-45-6789, diagnosed with hypertension.",
    context={
        "tenant_id": "hospital-001",
        "case_id":   "patient-123",
        "thread_id": "consultation-1",
        "actor_id":  "doctor-456",
    },
    llm_client=lambda prompt: f"Acknowledged for {prompt[:30]}..."
)

print(result.sanitized_text)  # Patient PERSON_1, SSN_REDACTED, diagnosed with hypertension.
print(result.final_text)      # Patient John Doe, SSN 123-45-6789, diagnosed with hypertension.

Note: The package is published as pii-firewall on PyPI; the import namespace is privacy_firewall.

---

Why PII Firewall?

---

Quick start with real detection (Presidio)

pip install "pii-firewall[presidio,langdetect]"
python -m spacy download en_core_web_sm

from privacy_firewall import create_firewall

firewall = create_firewall("finance", detector_backend="presidio")

result = firewall.secure_call(
    text="Card ending 4242 belongs to john@example.com.",
    context={"tenant_id": "acme", "case_id": "c1", "thread_id": "t1", "actor_id": "u1"},
    llm_client=lambda prompt: f"Processed: {prompt}"
)

print(result.sanitized_text)  # Card ending MASKED_4242 belongs to EMAIL_1.
print(result.final_text)      # Card ending 4242 belongs to john@example.com.

Pick your domain profile

---

Real-world use cases

Healthcare chatbot

A patient asks: "I'm John Doe, born 1985-03-12, my SSN is 123-45-6789."

The LLM receives: "I'm PERSON_1, born YEAR_1980_1989, my SSN is REDACTED."

The response is rehydrated before the user sees it.

Fintech support automation

A support ticket contains IBANs, card numbers, and email addresses. The LLM summarizes the issue without ever seeing real account data.

Legal document review

Party names, dates, and locations are pseudonymized so teams can run contract analysis through LLMs without exposing client data.

---

Detection backends

---

Documentation & links

• Full documentation
• Interactive demo (Colab)
• Website
• PyPI package

---

License

Apache 2.0. See LICENSE.