Skip to main content

PII Sanitization & Privacy Protection

AdMesh SDKs include built-in Personally Identifiable Information (PII) sanitization to ensure user privacy while maintaining recommendation quality. This feature automatically detects and removes sensitive information before sending queries to our API.

🔒 Privacy-First Design​

Core Principles​

  • Local Processing: All PII detection happens on your device/server
  • Zero External Calls: No data sent to third parties during sanitization
  • Complete Removal: PII is fully stripped from API requests
  • No Storage: Original input is never logged or stored
  • Standards Compliance: Meets GDPR, CCPA, and SOC 2 requirements

What Gets Detected & Removed​

  • Names: "I'm John", "My name is Sarah", "This is Alice"
  • Email Addresses: All standard formats including complex domains
  • Phone Numbers: US/international formats with various separators
  • Contextual Information: Age, gender, goals (extracted but not sent as PII)

🚀 Quick Start​

Python SDK​

from admesh import sanitize_and_build

# Raw user input with PII
user_input = "Hi, I'm Priya (priya@gmail.com). I'm a 27-year-old female building a wellness app."

# Sanitize and build clean prompt
result = sanitize_and_build(user_input)

print(result["prompt"])
# Output: "Suggest tools for a 27-year-old female building a wellness app."

print(result["removed"])
# Output: {"name": "Priya", "email": "priya@gmail.com", "phone": None}

TypeScript SDK​

import { sanitizeAndBuild } from 'admesh';

const userInput = "Hi, I'm Priya (priya@gmail.com). I'm a 27-year-old female building a wellness app.";
const result = sanitizeAndBuild(userInput);

console.log(result.prompt);
// Output: "Suggest tools for a 27-year-old female building a wellness app."

console.log(result.removed);
// Output: {name: "Priya", email: "priya@gmail.com", phone: null}

📝 Example Prompts for Testing​

Try these example prompts to see the sanitization in action:

Complete Scenario (All PII Types)​

Input: "Hello! I'm Sarah Johnson (sarah.j@techcorp.com, 555-123-4567). I'm a 32-year-old woman building a fintech startup focused on small business lending."

Expected Output:
- Prompt: "Suggest tools for a 32-year-old female building a fintech startup focused on small business lending."
- Removed: {name: "Sarah Johnson", email: "sarah.j@techcorp.com", phone: "555-123-4567"}
- Context: {age: 32, gender: "female", goal: "building a fintech startup focused on small business lending"}

Age + Goal Only​

Input: "I'm 28 years old and working on creating a mobile app for fitness tracking and nutrition planning."

Expected Output:
- Prompt: "Suggest tools for someone working on creating a mobile app for fitness tracking and nutrition planning (age 28)."
- Removed: {name: null, email: null, phone: null}
- Context: {age: 28, gender: null, goal: "working on creating a mobile app for fitness tracking and nutrition planning"}

Goal Only​

Input: "Looking for tools to help with building an e-commerce platform for sustainable fashion brands."

Expected Output:
- Prompt: "Suggest tools for building an e-commerce platform for sustainable fashion brands."
- Removed: {name: null, email: null, phone: null}
- Context: {age: null, gender: null, goal: "building an e-commerce platform for sustainable fashion brands"}

Contact Info Only (No Context)​

Input: "You can reach me at alex.developer@gmail.com or call (555) 987-6543 for more details."

Expected Output:
- Prompt: "Suggest relevant tools and services."
- Removed: {name: null, email: "alex.developer@gmail.com", phone: "(555) 987-6543"}
- Context: {age: null, gender: null, goal: null}

Edge Cases​

Input: "I'm Mike! 🚀 Building a crypto-trading bot... Age: 25!!! Contact: mike@crypto.io"

Expected Output:
- Prompt: "Suggest tools for a 25-year-old male building a crypto-trading bot."
- Removed: {name: "Mike", email: "mike@crypto.io", phone: null}
- Context: {age: 25, gender: "male", goal: "building a crypto-trading bot"}

🔧 Integration with Recommendations​

Python Integration​

import os
from admesh import Admesh, sanitize_and_build

client = Admesh(api_key=os.environ.get("ADMESH_API_KEY"))

def get_safe_recommendations(user_input: str):
    # Sanitize user input
    sanitized = sanitize_and_build(user_input)
    
    # Use clean prompt for recommendations
    response = client.recommend.get_recommendations(
        query=sanitized["prompt"],
        format="auto"
    )
    
    # Log what was removed (for debugging, not storage)
    print(f"PII removed: {sanitized['removed']}")
    
    return response

TypeScript Integration​

import Admesh, { sanitizeAndBuild } from 'admesh';

const client = new Admesh({
  apiKey: process.env.ADMESH_API_KEY
});

async function getSafeRecommendations(userInput: string) {
  // Sanitize user input
  const sanitized = sanitizeAndBuild(userInput);
  
  // Use clean prompt for recommendations
  const response = await client.recommend.getRecommendations({
    query: sanitized.prompt,
    format: 'auto'
  });
  
  // Log what was removed (for debugging, not storage)
  console.log('PII removed:', sanitized.removed);
  
  return response;
}

📊 Performance & Accuracy​

Performance Metrics​

  • Processing Time: less than 100ms for typical inputs
  • Memory Usage: Minimal (pre-compiled patterns)
  • Network Impact: Zero (local processing only)
  • Thread Safety: Fully thread-safe in all languages

Accuracy Metrics​

  • Name Detection: 95%+ accuracy across common patterns
  • Email Detection: 99%+ accuracy for standard formats
  • Phone Detection: 95%+ accuracy for US/international formats
  • False Positives: less than 2% (common words filtered out)

🛡️ Security Standards Compliance​

GDPR Compliance​

  • ✅ Data Minimization: Only necessary context extracted
  • ✅ Purpose Limitation: PII used only for sanitization
  • ✅ Storage Limitation: No PII storage or logging
  • ✅ Privacy by Design: Built-in privacy protection

CCPA Compliance​

  • ✅ Transparency: Clear documentation of data handling
  • ✅ User Control: Developers control what gets processed
  • ✅ No Sale: PII never shared or sold
  • ✅ Deletion: PII immediately discarded after processing

SOC 2 Type II​

  • ✅ Security: Local processing prevents data exposure
  • ✅ Availability: No external dependencies
  • ✅ Processing Integrity: Consistent, reliable sanitization
  • ✅ Confidentiality: PII never leaves your environment

🔍 Advanced Usage​

Custom Sanitization Workflows​

from admesh.sanitizer import PIISanitizer
from admesh.builder import PromptBuilder

# Use components separately for custom workflows
sanitizer = PIISanitizer()
builder = PromptBuilder()

# Analyze text
analysis = sanitizer.analyze_text("I'm Sarah, building a mobile app")

# Custom prompt building
if analysis['extracted_context']['age']:
    prompt = f"Age-specific tools for {analysis['extracted_context']['goal']}"
else:
    prompt = builder.build_complete_prompt(
        analysis['sanitized_text'], 
        analysis['extracted_context']
    )

Validation & Testing​

def validate_sanitization(test_input: str, expected_pii: dict):
    """Validate that PII sanitization works as expected."""
    result = sanitize_and_build(test_input)
    
    # Check that expected PII was detected
    for pii_type, expected_value in expected_pii.items():
        assert result["removed"][pii_type] == expected_value
    
    # Ensure no PII remains in prompt
    for pii_value in expected_pii.values():
        if pii_value:
            assert pii_value not in result["prompt"]
    
    return True

# Test with known inputs
validate_sanitization(
    "I'm John (john@test.com)",
    {"name": "John", "email": "john@test.com", "phone": None}
)

📞 Support & Questions​

For questions about PII sanitization:

🔄 Updates & Changelog​

Version 1.0.0 (Current)​

  • ✅ Initial PII sanitization release
  • ✅ Support for names, emails, phone numbers
  • ✅ Context extraction (age, gender, goals)
  • ✅ Python and TypeScript SDKs
  • ✅ Comprehensive test coverage

Planned Features​

  • 🔄 Additional PII types (addresses, SSNs)
  • 🔄 Custom pattern support
  • 🔄 Multilingual support
  • 🔄 Enhanced context extraction