redact-proxy-reid 0.1.2

DE-ID/RE-ID SDK for LLMs - Full-cycle PHI protection with automatic re-identification

Redact Proxy RE-ID SDK

Full-cycle PHI protection for LLMs with automatic re-identification.

Unlike simple PHI redaction which permanently removes sensitive data, this SDK:

1. Tokenizes PHI with unique reversible tokens (John Smith → [NAME_a1b2c3])
2. Sends tokenized text to your LLM (OpenAI, Anthropic, Gemini)
3. Re-identifies the response by restoring original PHI values

Your PHI never leaves your environment. LLMs only see anonymized tokens.

Installation

pip install redact-proxy-reid

# With specific LLM support
pip install redact-proxy-reid[openai]
pip install redact-proxy-reid[anthropic]
pip install redact-proxy-reid[all]  # All LLM providers

Quick Start (2 minutes)

Get Your API Key

1. Sign up at redact.health
2. Go to Dashboard → API Keys
3. Create a new RE-ID API key (starts with rr_live_)

Basic Usage

from redact_proxy_reid import PHITokenizer, PHIReidentifier

# Set your API key (or use REDACT_API_KEY env var)
API_KEY = "rr_live_your_key_here"

# 1. Tokenize PHI
tokenizer = PHITokenizer(api_key=API_KEY)
result = tokenizer.tokenize("Patient John Smith, DOB 01/15/1980, SSN 123-45-6789")

print(result.tokenized_text)
# "Patient [NAME_a1b2c3], DOB [DATE_d4e5f6], SSN [SSN_g7h8i9]"

# 2. Send to your LLM (tokens are safe!)
# llm_response = your_llm_call(result.tokenized_text)

# 3. Re-identify the response
reidentifier = PHIReidentifier(api_key=API_KEY)
restored = reidentifier.reidentify(llm_response, result.token_map)

print(restored.text)
# Original PHI values restored

Drop-in OpenAI Wrapper

from openai import OpenAI
from redact_proxy_reid import OpenAIWrapper

# Wrap your existing client
client = OpenAI()
wrapped = OpenAIWrapper(client, api_key="rr_live_your_key_here")
# Or set REDACT_API_KEY env var and omit api_key

# Use exactly like normal - PHI protection is automatic
response = wrapped.chat(
    model="gpt-4",
    messages=[{
        "role": "user",
        "content": "Summarize this patient's condition: John Smith, 45yo male, MRN 12345"
    }]
)

print(response["content"])
# Response with original PHI restored automatically

Drop-in Anthropic Wrapper

from anthropic import Anthropic
from redact_proxy_reid import AnthropicWrapper

client = Anthropic()
wrapped = AnthropicWrapper(client, api_key="rr_live_your_key_here")

response = wrapped.message(
    model="claude-3-opus-20240229",
    messages=[{
        "role": "user",
        "content": "Patient Jane Doe needs a referral for her diabetes management"
    }]
)

Drop-in Gemini Wrapper

import google.generativeai as genai
from redact_proxy_reid import GeminiWrapper

genai.configure(api_key="your-gemini-key")
model = genai.GenerativeModel("gemini-pro")
wrapped = GeminiWrapper(model, api_key="rr_live_your_key_here")

response = wrapped.generate(
    "Create a care plan for patient Bob Johnson, age 67"
)

Using Environment Variables

# Set once, use everywhere
export REDACT_API_KEY="rr_live_your_key_here"

# No api_key parameter needed - uses env var automatically
tokenizer = PHITokenizer()
wrapped = OpenAIWrapper(client)

Linking to Email (Optional)

You can link your API key to an email for account recovery and notifications:

tokenizer = PHITokenizer(
    api_key="rr_live_your_key_here",
    email="[e-mail]"
)

Multi-turn Conversations

Token mappings persist across conversation turns:

wrapped = OpenAIWrapper(client)

# First message
response1 = wrapped.chat(
    model="gpt-4",
    messages=[{"role": "user", "content": "Patient John Smith has diabetes"}],
    conversation_id="conv-123"
)

# Second message - same conversation
response2 = wrapped.chat(
    model="gpt-4",
    messages=[
        {"role": "user", "content": "Patient John Smith has diabetes"},
        {"role": "assistant", "content": response1["tokenized_content"]},
        {"role": "user", "content": "What medications should John take?"}
    ],
    conversation_id="conv-123"
)

# "John Smith" is consistently tokenized across both turns

Tier Configuration

Different tiers offer different capabilities:

from redact_proxy_reid import PHITokenizer, TierConfig, PHIType, TokenFormat

# Basic tier - sensible defaults
tokenizer = PHITokenizer(config=TierConfig.basic())

# Pro tier - more customization
config = TierConfig.pro()
config.tokenize_types = [PHIType.NAME, PHIType.SSN, PHIType.MRN]  # Only these types
config.token_format = TokenFormat.ANGLE  # <NAME_a1b2c3> instead of [NAME_a1b2c3]
tokenizer = PHITokenizer(config=config)

# Enterprise tier - full control
config = TierConfig.enterprise()
config.token_id_length = 10  # Longer, more unique tokens
config.custom_format = "<<{type}:{id}>>"  # Custom format
tokenizer = PHITokenizer(config=config)

PHI Types Supported

• NAME - Patient and provider names
• DATE - Dates of birth, admission, discharge
• SSN - Social Security Numbers
• PHONE - Phone numbers
• EMAIL - Email addresses
• ADDRESS - Street addresses
• MRN - Medical Record Numbers
• FACILITY - Hospital/clinic names
• AGE - Ages over 89
• ZIP - ZIP codes
• ACCOUNT - Account numbers, Medicare/Medicaid IDs
• LICENSE - License plate numbers
• VIN - Vehicle identification numbers
• DEVICE - Device identifiers
• URL - Web URLs
• IP - IP addresses

Token Map Serialization

Save and restore token maps for later re-identification:

import json

# Tokenize
result = tokenizer.tokenize("Patient data here")

# Save token map
token_map_json = json.dumps(result.token_map.to_dict())
# Store securely (database, encrypted file, etc.)

# Later: restore and re-identify
from redact_proxy_reid import TokenMap

token_map = TokenMap.from_dict(json.loads(token_map_json))
restored = reidentifier.reidentify(llm_response, token_map)

Selective Re-identification

Re-identify only certain PHI types:

# Only restore names, keep dates tokenized
restored = reidentifier.reidentify(
    text=llm_response,
    token_map=token_map,
    types_to_restore=[PHIType.NAME]
)

How It Works

Your App                    Redact RE-ID SDK                    LLM Provider
    |                              |                                  |
    |  "Patient John Smith..."     |                                  |
    | ---------------------------> |                                  |
    |                              |                                  |
    |                    Tokenize: "Patient [NAME_a1b2c3]..."         |
    |                    Store: {[NAME_a1b2c3]: "John Smith"}         |
    |                              |                                  |
    |                              |  "[NAME_a1b2c3]..."              |
    |                              | -------------------------------> |
    |                              |                                  |
    |                              |  "...[NAME_a1b2c3]..."           |
    |                              | <------------------------------- |
    |                              |                                  |
    |                    Re-identify: "...John Smith..."              |
    |                              |                                  |
    |  "...John Smith..."          |                                  |
    | <--------------------------- |                                  |

Security Notes

• PHI tokenization happens locally in your environment
• Only anonymized tokens are sent to LLM providers
• Token maps should be stored securely (they contain the PHI!)
• Consider encrypting token maps at rest

License

Commercial license required. See redact.health for pricing.

Links

• Product Page
• Documentation
• Support