Blind AI, Seeing Reader: Local Document Sanitization for LLMs

This is a problem familiar to anyone who has connected an LLM to work documents: the model performs better when it sees the full context, but it is precisely this context that most often contains information that cannot be shared externally. Provexar’s Sanitizer offers a specific architectural solution—processing documents locally before they reach the AI.

The main idea is reversible sanitization. Names, contacts, credentials, and keys aren’t permanently deleted but are replaced with opaque tokens according to the project’s policy. The AI receives text that contains no real identifiers, processes it, and returns a response. Then, the tokens are reversed locally—you see the result with the real names, and the model never knew them. This is a reasonable compromise: you don’t sacrifice accuracy for security, nor do you sacrifice security for accuracy.

The second layer is dealing with formatting clutter. A 40 MB PDF or Word file consists mostly of layout, embedded fonts, and metadata. The Sanitizer converts everything to clean Markdown and flat CSV, saving tokens and context slots for the actual content rather than binary bloat. For those who track the cost of LLM queries, this isn’t just a cosmetic improvement—it’s real money.

The third component is indexing. Each document is chunked and indexed twice: a full-text index for exact terms and a semantic index for meaning-based search. AI agents receive extraction tools and pull five relevant paragraphs instead of five hundred pages, citing the source. Essentially, Sanitizer builds a local RAG infrastructure, but with an emphasis on sanitization as the first step in the pipeline.

Everything runs locally: conversion, OCR, sanitization, indexing, and search. No cloud services or downloads. This addresses the main concern for corporate use—that data will leak through the provider’s API.

The restrictions are fair, too. The personal version is free, but the installer isn’t digitally signed—Windows SmartScreen will display a warning, and you’ll need to manually verify the SHA-256 hash. The business version, priced at $49.99, is distributed through the Microsoft Store and is signed by Microsoft. Windows 10/11, 64-bit only. The tool is not yet available for Linux or macOS.

The license is straightforward: the personal version is for non-commercial use only; everything else requires the Business edition. Reverse engineering is prohibited; no warranties are provided.

Reversible sanitization is not a panacea. Tokenization works when sensitive data can be identified using patterns or a dictionary. But if confidential information is scattered throughout the text in free-form, pattern-based replacement won’t catch it. The tool mitigates most of the risk, but the responsibility for verifying the sanitized output remains with the user—this is explicitly stated in the license.

Sanitizer is interesting not as just another AI product, but as a practical layer between your documents and a third-party model. The question to ask is: to what extent does tokenization cover your actual types of sensitive data before you trust it with production documents?

Source: Hacker News - Newest: ""AI" "LLM""