A Taxonomy and Survey of Multi-Agent Retrieval- Augmented Generation: Architectures, Evaluation, and Open Challenges

Retrieval-Augmented Generation (RAG) was an addition created to support large language model (LLM) outputs in external, verifiable knowledge—without needing to spend so much money to reduce hallucination. model retraining. However, single-agent RAG pipelines fall short every time when tasks require multi-hop. reasoning, noisy retrieved content or heterogeneous retrieved content or domain specific disambiguation. Multi-Agent RAG (MARAG) tackles these limitations by breaking down the retrieval-generation process into distinct tasks. Pipeline across an ensemble of specialized agents: Planners, retrievers, filters, reasoners, Each of these provides a specific cognitive function: synthesisers, and verifiers. This paper offers a In-depth survey of the MA-RAG landscape. The field of development is traced back to its beginnings and Take the systems from reflective to full collaborative architecture, and add a taxonomy of seven categories. Analyse typical examples of each system. Communication protocols, memory designs, The following reasoning strategies and verification mechanisms are explored in detail: benchmark coverage, and evaluation metrics. We conclude by stating seven open challenges and a list of high-priority Directions that include adaptive topologies, scalable retrieval, standardised evaluation, and trustworthy. deployment.