LexiRedact: A Middleware Architecture for Privacy-Utility Balanced Document Redaction Using Dual Path Processing

Varad Limbkar , Baihela Hussain , Shwetan Londhe , Vijay Rathi

Data Privacy, PII Detection, Privacy-Preserving AI, Retrieval-Augmented Generation, Vector Databases.

As Retrieval-Augmented Generation (RAG) systems become integral to enterprise AI workflows, the persistence of sensitive textual data within vector databases introduces critical privacy and compliance risks. Existing mitigation approaches force a tradeoff between privacy and utility: pre-redaction techniques significantly degrade semantic retrieval accuracy, while encryption-based controls violate data minimization principles and increase operational complexity. This work presents LexiRedact, an innovative, ML assisted middleware framework that addresses this tradeoff through a dual-path processing architecture. The framework decouples the ingestion pipeline into parallel streams: a utility path generates high-fidelity semantic embeddings from original content to preserve contextual accuracy, while a privacy path performs concurrent PII detection and sanitization prior to persistent storage using pre-trained machine learning components. Machine learning components used for PII detection and embedding generation are incorporated as modular elements informed by established evaluations reported in prior work, enabling the framework to focus on architectural design and system level optimization. To mitigate the latency overhead of real time interception, system level optimizations including asynchronous execution, batch processing, and Redis assisted semantic caching are employed. Experimental evaluation based on systematic latency measurements demonstrates that the architecture achieves sub-150ms ingestion latency for cached operations and improves overall system throughput by approximately 30 to 40 percent compared to synchronous baselines. Overall, LexiRedact provides a scalable, privacy-preserving ingestion infrastructure that enforces data minimization while maintaining the operational efficiency of RAG systems.

Volume 14 Issue 3, March-2026

Pages : i895-i905