Privacy-Preserving Structured Knowledge Extraction from Census-Style Records Using a Hierarchical Multi-Agent Open-Weight LLM Architecture

Abstract

Census-style address records contain heterogeneous structures that must be decomposed into fine-grained fields before they can support linkage, geocoding, and administrative processing. This paper describes the design and implementation of a three-stage Planner–Manager–Worker pipeline using a locally deployed open-weight language model. The prototype includes address-specific extraction, same-model semantic consistency review, formatting, record-preserving chunking, and an exploratory bounded retry pathway. Although the prototype also implements workers for names, phone numbers, email addresses, Social Security numbers, and dates of birth, the quantitative evaluation is limited to address-component parsing. We conduct a controlled pilot using 700 author-generated synthetic records distributed across seven address categories. Each record was first created as a structured object and then rendered into text; the original structured object served as the evaluation reference. Across three repeated executions on the same fixed benchmark, the complete hierarchical configuration obtained a mean micro-averaged address-component exact-match score of 95.7%, compared with 80.9% for the evaluated monolithic LLM reference; the deterministic rule-based reference obtained 52.3%. The fixed prompts were not matched, and realized inference calls, dynamic token allocation, original batching, and reported-run segmentation could not be verified from retained artifacts. No component ablations were conducted. Consequently, the numerical difference cannot be attributed uniquely to planning, task coordination, specialization, semantic review, feedback, chunking, or multi-agent organization. The reference labels were not independently annotated, the benchmark does not estimate performance on naturally occurring records, and the retry pathway was invoked only three times. The contribution is therefore the specification and controlled pilot characterization of an on-premise address-parsing pipeline, rather than evidence of architectural causality, production readiness, broad PII-extraction accuracy, or real-record robustness.