AN ENTERPRISE MULTI-AGENT AI FRAMEWORK FOR HR DOCUMENT INTELLIGENCE USING RETRIEVAL-AUGMENTED GENERATION AND VECTOR SEARCH

Abstract

Human Resource (HR) departments in universities and large enterprises remain constrained by manual document workflows: policy retrieval, compliance verification, and correspondence drafting are performed by administrators reading dense unstructured charters page by page. Existing cloud-based Large Language Model (LLM) solutions promise automation but introduce three unacceptable enterprise risks: factual hallucination, opaque “black-box” reasoning, and cross-border transmission of personally identifiable information. This paper proposes an Enterprise Multi-Agent HR Document System that resolves these limitations through a local-first cognitive architecture combining Multi-Agent Systems (MAS), Retrieval-Augmented Generation (RAG), and high-performance vector indexing. A Chief Orchestrator agent delegates sub-tasks to three specialized workers (a Document Finder, a Compliance Vet, and a Document Writer), each governed by a bounded system prompt and an isolated toolset. Institutional documents are ingested through a sliding-window chunking pipeline (chunk size c = 1000, overlap o = 200), encoded into 384-dimensional dense vectors using the all-MiniLM-L6-v2 Sentence-Transformer, and indexed in a FAISS store paired with a structured JSON metadata registry that preserves the source file path, section heading, page number, and version tag of every chunk. Quantized LLMs (Llama-3-8B-Instruct) are served locally via Ollama, keeping all sensitive HR data on-premise. Empirical evaluation on 150 pages of university bylaws (≈450 chunks) shows sub-second FAISS retrieval (2.1 ms per query on the optimized workstation), a Precision@3 of 0.942 on synonym-heavy natural-language queries, and an approximately 90% reduction in end-to-end document retrieval latency relative to manual administrative lookup. Every generated response is traceable through the JSON registry back to its exact source paragraph, closing the enterprise auditability gap that has historically blocked adoption of generative AI in regulated HR workflows.