llm-redact 0.1.1

Privacy-first text redaction using local LLM models with rule generation capabilities

LLM Redact

Privacy-first text redaction using local LLM models. Automatically detect and redact sensitive information like names, emails, phone numbers, and more.

Features

• 🔒 Privacy-first - Uses local LLM models, no data sent to external services
• 🚀 Simple API - One-liner redaction: llm_redact.mask(text)
• 💾 Smart Caching - SQLite database for caching and history
• 🔧 Configurable - Custom rules, models, and database connections
• 📊 Tracking - Full history and analytics of redaction operations

Installation

pip install llm-redact

Quick Start

import llm_redact

# Simple redaction
result = llm_redact.mask("Hi, I'm John Doe from [e-mail]")
print(result.redacted_text)
# Output: "Hi, I'm |_NAME_A1B2C3D4_| from |_EMAIL_E5F6G7H8_|"

print(result.replacements)
# Output: [
#   Replacement(original_text="John Doe", replacement_text="|_NAME_A1B2C3D4_|"),
#   Replacement(original_text="[e-mail]", replacement_text="|_EMAIL_E5F6G7H8_|")
# ]

# Note: Placeholders contain unique IDs and can be stored in database for restoration
# Each placeholder like |_NAME_A1B2C3D4_| maps to original text via database lookup

Configuration

Environment Variables

# LLM Host (default: http://localhost:8000)
export LLM_REDACT_LLM_HOST_URL=http://localhost:8000

# Database (default: sqlite:///llm_redact.db)
export LLM_REDACT_DATABASE_URL=sqlite:///my_redact.db

# Model (default: gemma3:1b)
export LLM_REDACT_DEFAULT_MODEL=gemma3:1b

# Caching (default: True)
export LLM_REDACT_ENABLE_CACHING=true

Custom Database

import llm_redact

# Use PostgreSQL
llm_redact.configure_client(
    database_url="postgresql://user:pass@localhost/redact_db"
)

# Use custom LLM host
llm_redact.configure_client(
    llm_host_url="http://my-llm-server:8000"
)

Advanced Usage

Custom Rules

from llm_redact import RedactionRule

custom_rules = [
    RedactionRule(
        name="Replace SSN with [SSN]", 
        description="Social Security Numbers",
        data_type="SSN"
    ),
    RedactionRule(
        name="Replace addresses with [ADDRESS]", 
        description="Physical addresses",
        data_type="ADDRESS"
    )
]

result = llm_redact.mask(
    "My SSN is 123-45-6789 and I live at 123 Main St",
    rules=custom_rules
)

Using the Client Directly

from llm_redact import LLMRedactClient

client = LLMRedactClient(
    llm_host_url="http://localhost:8000",
    database_url="sqlite:///custom.db"
)

result = client.mask("Sensitive text here")

# Get history
history = client.get_history(limit=10)

# Create custom rules
rule = client.create_rule(
    name="Replace API keys with [API_KEY]",
    description="API keys and tokens"
)

Prerequisites

1. LLM Host Server: Run the llm-redact host server locally:

   # Install and run the LLM host
   ollama serve
   ollama pull gemma3:1b
   
   # Run llm-redact host server
   python -m llm_redact_host
   

2. Database: SQLite (default) or any SQLAlchemy-supported database

Supported Redaction Types

• Personal names → |_NAME_XXXX_|
• Email addresses → |_EMAIL_XXXX_|
• Phone numbers → |_PHONE_XXXX_|
• Countries → |_COUNTRY_XXXX_|
• Universities → |_UNIVERSITY_XXXX_|
• Job titles → |_JOB_TITLE_XXXX_|
• Addresses → |_ADDRESS_XXXX_|
• Social Security Numbers → |_SSN_XXXX_|
• Credit card numbers → |_CREDIT_CARD_XXXX_|

Where XXXX is a unique 8-character hash ID for each piece of data.

API Reference

llm_redact.mask(text, rules=None, model=None)

Redact sensitive information from text.

Parameters:

• text (str): Text to redact
• rules (list, optional): Custom redaction rules
• model (str, optional): LLM model to use

Returns: RedactionResult object

RedactionResult

• original_text: Original input text
• redacted_text: Text with sensitive data redacted
• replacements: List of replacements made
• is_redacted: Whether any redactions were made
• processing_time_ms: Processing time in milliseconds
• cached: Whether result was from cache

License

MIT License