5 |
6 | This repository provides an **Electron** desktop application for **locally anonymizing documents** before sending them to advanced Large Language Models (LLMs). By stripping out personal or identifiable information using a **context-aware** model, you can safely train or query external LLMs (e.g., OpenAI’s o3 model) with minimal privacy risk.
7 |
8 | ## Motivation
9 |
10 | - **PII Removal**: Traditional RegEx-based anonymization often fails on nuanced data. With an ONNX-based model, you gain context-aware detection for **names, addresses, phone numbers, etc.**
11 | - **Safe LLM Usage**: Many companies need to keep real customer or employee data **internal** but still want to leverage powerful external LLMs. This tool helps them do so by anonymizing data **on their end** first.
12 | - **Flexible**: Supports `.txt`, `.docx`, `.xls(x)`, `.csv`, `.pdf`, and more. Converts text, merges tokens, and replaces them with consistent pseudonyms.
13 |
14 | ## Key Features
15 |
16 | 1. **Electron App**: Cross-platform desktop UI built on HTML/CSS/JS.
17 | 2. **Daily Limit**: **100** documents per day for the free tier (by default). You can raise or remove it if you prefer—this is open source.
18 | 3. **Context-Aware**: Relies on a local ONNX model downloaded separately (due to GitHub’s file-size constraints).
19 | 4. **Mapping** (Pro Mode): If you enable Pro, the app can produce a JSON file mapping each original entity (e.g., “John Smith”) to its anonymized token (e.g., “NAME_1”).
20 | 5. **MIT License**: Free to modify and distribute. We welcome contributions.
21 |
22 | ## Getting Started
23 |
24 | 1. **Clone or Download** this repository.
25 | - Model available here: [https://huggingface.co/iiiorg/piiranha-v1-detect-personal-information/tree/main](https://huggingface.co/iiiorg/piiranha-v1-detect-personal-information/tree/main)
26 | - Download the ONNX model (`.onnx` files) from our external link (not included here due to size constraints).
27 | - Place it under `./models/protectai/lakshyakh93-deberta_finetuned_pii-onnx/` or as directed in `fileProcessor.js`.
28 | 3. **Install Dependencies**:
29 | ```bash
30 | npm install
31 | ```
32 | 4. **Run (Dev Mode)**:
33 | ```bash
34 | npm run dev
35 | ```
36 | Or if you prefer:
37 | ```bash
38 | npx electron .
39 | ```
40 | 5. **Build / Package** (macOS example):
41 | ```bash
42 | npm run build:mac
43 |
44 |
45 | ### Basic Usage
46 |
47 | - **Drop or Select Files**: The main UI allows you to drag-and-drop or pick multiple files/folders.
48 | - **Output Directory**: Choose where the anonymized files should be placed.
49 | - **Anonymize**: Click “Anonymize Files” to run.
50 | - **Mapping (Pro)**: If you have a Pro key, the app will create an additional `-map.json` file capturing each replaced entity.
51 |
52 | ## How It Works
53 |
54 | - **Electron**:
55 | - **`main.js`**: Spawns the main window, handles file selection, passes tasks to `FileProcessor`.
56 | - **`renderer.js`**: Manages the UI (index.html), user interactions, daily usage counters, and “Pro” logic.
57 | - **`fileProcessor.js`**:
58 | - Loads the local ONNX model (via `@xenova/transformers`).
59 | - Identifies personal data by context (names, addresses, etc.).
60 | - Replaces them with tokens (`NAME_1`, `PHONE_NUMBER_3`, etc.).
61 | - If Pro, writes a JSON mapping for re-identification.
62 | - **Local Model**:
63 | - We rely on a context-aware token classification model. This is significantly more effective than simple RegEx for real-world PII.
64 |
65 | ## Limitations & Notes
66 |
67 | - **Daily 100-File Limit**: By default, the free version only processes 100 documents per day. This is purely enforced in the UI. Since it’s open source, you can remove or change it as needed.
68 | - **No Guarantee**: Even context-aware models can miss certain edge cases. Always manually review if 100% privacy is critical.
69 | - **Cross-Platform**:
70 | - macOS builds are tested on both M-series (ARM) and Intel.
71 | - Windows and Linux builds are also available.
72 |
73 | ## Contributing
74 |
75 | We use the **MIT License**, so feel free to open pull requests, modify the code, or adapt it for your needs. If you make improvements or fix bugs, please share them back!
76 |
77 | ---
78 |
79 | **Thanks for checking out the A5 PII Anonymizer.** We hope this helps you safely leverage powerful LLMs with real data while keeping personal information private.
--------------------------------------------------------------------------------
/index.html:
--------------------------------------------------------------------------------
1 |
2 |
3 |
4 |
21 |
22 | 
23 | PII Anonymizer
24 |
25 |
26 |
27 |
28 |
29 |
38 |
39 |
40 | Drag & Drop files here or click to select
30 | 37 |
41 |
44 |
45 |
46 |
47 |
53 |
54 |
55 |
56 |
66 |
67 |
68 |
71 |
72 |
73 |
76 |
77 |
78 |
79 |
80 |
81 |
85 |
86 |
87 |
90 |
57 |
58 |
65 |
59 |
60 |
63 |
64 |