Multi-Agentic Retriever (MAR): Redefining Information Retriever in the Era of AI agents

Current information retrieval systems often rely on either one retriever or a rigid, pre-defined hybrid retrievers. These static, "one-size-fits-all" approaches lack the flexibility to adapt to the diverse nature of user queries, often leading to suboptimal performance on wide spectrum of queries from different datasets. This thesis introduces the Multi-Agentic Retriever (MAR), a flexible and autonomous framework that recasts information retrieval as a multi-agent problem. MAR utilizes specialized agents, including a \textbf{\textit{PlannerAgent}} for dynamic plan generation and a \textbf{\textit{RetrieverAgent}} equipped with a diverse toolkit of lexical (BM25), dense (Contriever), and sparse-neural (Splade) retrieval methods, alongside LLM-based filtering and reranking tools. We conduct a comprehensive series of ablation studies to systematically evaluate the performance, component interactions, and latency-accuracy trade-offs of this framework. Our key findings reveal a complex landscape. First, while static \textbf{Rule-Based plans} currently achieve the highest peak accuracy, the zero-shot \textbf{LLM-Generated plans} are "toe-to-toe" in performance. Counterintuitively, the LLM-Generated plans are also significantly \textbf{faster on average}, as the \textbf{\textit{PlannerAgent}} can dynamically select simpler, more efficient execution paths. Second, we identify a clear "sweet spot" for model scale: \textbf{32B models} consistently provide the best performance, as 7B/14B models lack the reasoning capabilities for complex planning, while 70B models introduce high latency for diminishing returns. Third, our analysis of agent tools uncovers a fundamental conflict: the \textbf{Query Expansion Agent} is a double-edged sword \textbf{strongly benefiting lexical retrievers} while simultaneously \textbf{harming dense and sparse-neural retrievers}. Finally, we demonstrate that the \textbf{\textit{LLMFilterTool}} is critical for achieving state-of-the-art performance, but its utility is entirely dependent on the reasoning capability of a large-scale model. This research concludes that the future of retrieval lies not in monolithic models but in flexible, multi-agent systems. The MAR framework serves as a robust foundation for this paradigm. We posit that with targeted fine-tuning, the \textbf{\textit{PlannerAgent}} can achieve full autonomy, and the MAR system, when integrated as the retrieval backbone for Retrieval-Augmented Generation (RAG) pipelines, can unlock a new generation of high-precision, intelligent information access systems.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.