Agentic Reterival Augmented Generation using Small Language Model

Abstract

Large language models (LLMs) have transformed artificial intelligence by enabling humanlike text generation and advanced natural language understanding. However, their dependence on static training data restricts their ability to answer evolving, real-time queries, often resulting in outdated or contextually inaccurate responses. Retrieval-Augmented Generation (RAG) addresses this limitation by integrating external knowledge sources, yet traditional RAG pipelines remain rigid, single-step, and unable to support adaptive reasoning or complex task execution, especially under computational constraints. Agentic Retrieval-Augmented Generation (Agentic RAG) with small language models (SLMs) overcomes these challenges by embedding autonomous agents within the retrieval and generation workflow. These agents incorporate design patterns such as planning, reflection, tool use, and multi-agent collaboration to dynamically refine queries, adjust retrieval strategies, and iteratively improve contextual grounding. Combined with domain-specific fine-tuning of SLMs on high-quality datasets, this architecture enables robust, scalable, and cost-efficient performance while maintaining real-time adaptability. This survey presents a comprehensive study of Agentic RAG using SLMs, tracing the evolution of RAG methods and detailing a taxonomy of agentic architectures. It further examines key use cases across healthcare, finance, legal systems, and education, alongside practical implementation strategies for production environments. Finally, it discusses open challenges related to scalability, operational reliability, ethical alignment, and performance optimization, offering insights into emerging frameworks and tools shaping the next generation of agentic RAG systems.

Ähnliche Arbeiten