Adaptive Decision Intelligence through Retrieval-Augmented Multi-Agent Large Language Models

Abstract

The integration of Retrieval-Augmented Generation (RAG) with multi-agent Large Language Models (LLMs) presents transformative potential for adaptive decision intelligence across critical domains. However, significant challenges including evaluation metric standardization, retrieval latency optimization, multi-agent consistency, dynamic knowledge evolution, and human-AI trust management hinder widespread adoption. This research proposes a comprehensive five-layer framework comprising foundation benchmarks, dynamic hybrid retrieval, multi-agent collaboration with weighted consensus, knowledge graph evolution through graph neural networks, and adaptive human-AI interaction. The system employs hybrid sparse-dense retrieval with reinforcement learning optimization, Byzantine fault-tolerant consensus protocols, and temporal knowledge graph embeddings for real-time updates. Experimental validation across healthcare, manufacturing, and cybersecurity domains demonstrates exceptional performance with 92.3% decision accuracy, 412ms retrieval latency, 94.7% agent agreement rate, and 1150 concurrent request capacity, exceeding all established thresholds. The framework effectively bridges the identified gaps, establishing a robust foundation for trustworthy, scalable, and truly adaptive decision intelligence systems with significant practical applicability across diverse domains.