International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 12 Issue: 04 | Apr 2025
p-ISSN: 2395-0072
www.irjet.net
From Retrieval to Reasoning: How Agentic AI Transforms Knowledge Retrieval and Content Generation Kunal Agarwal1 1AI/ML Computational Science Senior Analyst, Accenture
----------------------------------------------------------------------------***-------------------------------------------------------------------------Abstract This paper presents a novel agentic AI architecture that goes beyond traditional Retrieval-Augmented Generation (RAG) by embedding autonomous decision-making, goal-oriented planning, and self-reflection capabilities. Our implementation independently discovers, processes, and maintains its knowledge base while dynamically selecting tools and models based on task demands. Through continuous experiments with 180 queries spanning simple factual, creative, and complex analytical categories, we noticed that our agentic approach improved content relevance by 37% and user satisfaction by 42% compared to the conventional RAG systems. The agent's autonomous knowledge management reduced the manual intervention by 89% while maintaining freshness of crawled information. Perhaps most interestingly, the system's self-reflection mechanism enabled continuous learning, with satisfaction scores climbing by 14.2% during our four-week experimental period. These findings suggest that incorporating true agency into AI systems creates more adaptive, effective, and independent content generation capabilities—potentially reshaping how knowledge workers and AI systems collaborate. Key Words: Agentic AI, RAG, Autonomous, Artifical Intelligence, LLM, Vector Databse, Knowledge Graph, Pydantic, Deepseek, Llama
Introduction The history of artificial intelligence, from the very beginning to rule-based expert systems, through statistical learning approaches, to today's large language models (LLMs) that can generate text that is very close to human writing (Brown et al., 2020) has not been smooth. Despite the big achievements, the fact that most LLM systems are still merely reactive poses a hurdle. First introduced Traditional Retrieval-Augmented Generation (RAG) systems are an example of how this problem was tackled, whose major function was to pair the generated outputs with relevant information (Lewis et al., 2020; Guu et al., 2020). However, we've discovered that these systems usually still need human assistance in the process of knowledge acquisition and knowledge base maintenance -- a situation that leads to less independence and flexibility, particularly in cases of rapid changes in the knowledge domains (Shuster et al., 2022). 1.1 Literature Review Research on agentic AI has started long ago and can be taken back to Bratman (1987) who proposed the belief-desireintention agent models. This research line has moved through the recent stages of the development of the idea of an agent (Wooldridge, 2009). Along with the traditional features of agency in AI, the notion of the four isolated elements, that is, autonomy, social ability, reactivity, and proactiveness, has been ubiquitous in the literature since (Franklin & Graesser, 1996). It is undoubted that the development of language-based artificial intelligence solutions has been a target for many scholars. Weston et al. (2022) tried to embed the persistence and goal-oriented behavior in dialogue systems and their results were marvelous—yet the work merely resulted in a partial agency of the system. The same holds for Park et al. (2023) who further explained some memory mechanisms which in their case, the conversational agents displayed better user experience due to the memory mechanism. They have made significant progress toward coherent long-term interaction. Through the literature, WebGPT (Nakano et al., 2021) and REALM (Guu et al., 2020) are encountered as some of the systems that firstly utilized web searches yet the final analysis depicts that neither of them has all the necessary elements for the system to be considered fully agentic especially the capabilities to set goals and introspect autonomously (our work grows towards the latter two).
© 2025, IRJET
|
Impact Factor value: 8.315
|
ISO 9001:2008 Certified Journal
|
Page 180