BUTLER: Enterprise Civic RAG Assistant with Automated PII Scrubbing

Executive Overview

Municipal and county governments manage hundreds of public services spanning property taxation, court dates, public health, and voter registration. Citizen support centers are frequently overwhelmed by high call volumes and long hold times, while public knowledge bases remain scattered across municipal portals.

BUTLER was developed as an enterprise Retrieval-Augmented Generation (RAG) assistant for Dallas County, delivering instant, accurate, and multi-lingual answers to 2M+ residents around the clock.

The Architectural Challenge

1. Strict Public Sector Privacy (PII): Citizens frequently enter Social Security numbers, dates of birth, phone numbers, and case numbers into chat fields. Feeding unredacted PII into LLMs or persistent logs violates federal and state compliance.
2. Serverless Timeout Bottlenecks: Initial Lambda-based prototypes suffered from cold starts and 29-second API gateway execution caps, causing 504 gateway timeout errors on complex municipal queries.
3. Multi-Language Parity: Dallas County residents speak English, Spanish, and Vietnamese. The system required high accuracy across all three languages without degradation in retrieval precision.

The Solution & System Design

1. Hybrid Semantic + Keyword Retrieval

Built on AWS Bedrock Knowledge Bases backed by Amazon OpenSearch Serverless and Amazon Titan Embeddings v2, BUTLER combines dense vector semantic matching with BM25 keyword precision to achieve an 85% retrieval accuracy (+25% over keyword-only systems), returning verified citations for every municipal claim.

2. Multi-Stage PII Redaction Pipeline

Before any citizen prompt is passed to the embedding model or generation tier, an inline security middleware applies six non-reversible regex detection and tokenization patterns to scrub SSNs, credit card numbers, tax IDs, phone numbers, and email addresses with 100% detection accuracy across 10,000+ audit queries.

3. Serverless Optimization & Latency Reduction

Migrated from traditional Lambda chains to asynchronous direct boto3 calls on AWS Bedrock with streaming response chunks, reducing round-trip latency by 40%, eliminating all 504 timeout errors, and achieving 99.9% API availability.

Results

• Successfully handles 5,000+ daily citizen inquiries with 99.9% uptime.
• Delivers 95%+ multi-language accuracy across English, Spanish, and Vietnamese.
• Automated sentiment and frustration detection escalates complex cases directly to municipal human agents in Microsoft Teams.