MCP-Coordinated Multi-Agent RAG for Scalable Computational Biology and Mathematical Simulations

Abstract

We present a cloud-native framework that integrates a Model Context Protocol (MCP) server with multi-agent retrieval-augmented generation (RAG) and hybrid lexical–semantic retrieval (BM25 + FAISS) over multiple overlapping chunks. The system is orchestrated by agentic roles (Research, Reasoning, Validation, Orchestrator) and enforced via schema-constrained outputs to improve consistency, auditability, and robustness. We target mathematical biology tasks as a high-value testbed: epidemic modeling (SIR), gene regulatory and protein–protein interaction summaries, and hypothesis generation from primary literature. The architecture supports persistent context, cross-session continuity, and elastic scaling on public cloud. We outline formal components, end-to-end pipelines, and an evaluation protocol with benchmark- style metrics for factuality, grounding, and schema compliance. A compact case study demonstrates how hybrid retrieval and agentic orchestration reduce hallucinations and increase reproducibility for SIR analysis. We release this paper as a practical template for Q2-tier venues: fully reproducible diagrams, equations, and tables are provided for rapid adaptation.1