Network Inspector Guide

Verify NotPixel’s privacy claims yourself. This guide shows how to inspect network traffic and confirm that your users’ text never leaves their device when using privacy: true.

Quick Verification

What You’ll See

With privacy: true (Local Embeddings)

{
  "publisherId": "pub_abc123",
  "mode": "embedding",
  "embedding": [
    0.0234, -0.0891, 0.1456, 0.0023, -0.0567, 0.0891, ...
    // 384 floating-point numbers
  ],
  "context": {
    "lang": "en"
  }
}

What to notice:

• mode is "embedding" — indicates local embedding mode
• embedding contains only numbers — mathematical vector representation
• No input field — your text was never sent
• The 384 numbers are irreversible — cannot be converted back to text

With privacy: false (Standard Mode)

{
  "publisherId": "pub_abc123",
  "mode": "standard",
  "input": "help me optimize my SQL query for [email]",
  "context": {
    "lang": "en"
  }
}

What to notice:

• mode is "standard"
• input contains sanitized text — PII like emails are replaced with [email]
• Original email address is NOT visible

Step-by-Step: Chrome DevTools

Verification Checklist

Use this checklist to verify privacy mode is working correctly:

For privacy: true

For privacy: false (Standard)

Common Issues

”I see input field with my text”

Problem: Privacy mode is not enabled.

Solution: Ensure you have privacy: true in your SDK configuration:

const ads = new Ads({
  publisherId: 'pub_xxx',
  privacy: true,  // Must be explicitly set
});

“I see mode: "standard" instead of "embedding"”

Problem: Either privacy mode is off, or local embedding failed and fell back to standard mode.

Solution:

1. Check console for warnings about embedding failures
2. Ensure @huggingface/transformers is installed:

   npm install @huggingface/transformers

3. Preload the model to catch errors early:

   await ads.embedding.preload();
   console.log(ads.embedding.isReady()); // Should be true

“The embedding array has different length”

Problem: Different models produce different embedding dimensions.

Expected dimensions by model:

Automated Verification

You can also verify programmatically in your tests:

import Ads from 'notpixel';
 
// Mock fetch to inspect the request
const originalFetch = global.fetch;
global.fetch = async (url, options) => {
  if (url.includes('/v1/pick')) {
    const body = JSON.parse(options.body);
 
    // Verify privacy mode
    console.assert(body.mode === 'embedding', 'Should be embedding mode');
    console.assert(body.input === undefined, 'Should NOT have input field');
    console.assert(Array.isArray(body.embedding), 'Should have embedding array');
    console.assert(body.embedding.length === 384, 'Should have 384 dimensions');
 
    console.log('✅ Privacy verification passed!');
  }
  return originalFetch(url, options);
};
 
const ads = new Ads({
  publisherId: 'pub_xxx',
  privacy: true,
});
 
await ads.getAd({ input: 'my sensitive query' });

Understanding Embeddings

What is an embedding?

An embedding is a mathematical representation of text as a vector of numbers. It captures the semantic meaning without containing the actual words.

Text: "How do I optimize PostgreSQL queries?"
     ↓ (embedding model)
Vector: [0.023, -0.089, 0.145, 0.002, -0.056, ...]
        (384 floating-point numbers)

Why can’t embeddings be reversed?

1. Information loss: 384 numbers cannot encode all possible sentences
2. Many-to-one mapping: Different texts can produce similar embeddings
3. No dictionary: The model doesn’t store a word↔number mapping

Visual Example

┌─────────────────────────────────────────┐
│ "I need help with my database"          │
│ "Help me optimize my SQL"               │  Similar
│ "Database performance issues"           │  embeddings
└─────────────────────────────────────────┘
                    ↓
        [0.12, 0.34, -0.56, ...]  ← Close vectors
        [0.11, 0.35, -0.54, ...]
        [0.13, 0.33, -0.55, ...]

┌─────────────────────────────────────────┐
│ "Best pizza recipes"                     │  Different
└─────────────────────────────────────────┘  embedding
                    ↓
        [-0.45, 0.78, 0.12, ...]  ← Far vector

The ad matching system finds ads with similar embeddings, without ever knowing what the user actually typed.

Summary