LARGE LANGUAGE MODEL DRIVEN AGENTIC FRAME WORK FOR ADAPTIVE RETRIEVAL AND MULTI-HOP REASONING

Abstract

Retrieval Augmented Generation (RAG)hasestablished itself as a foundational paradigm for grounding the outputs of Large Language Models in externally retrieved factual ev idence, substantially reducing the hallucination that arises when parametric models are queried on knowledge beyond their training distribution. However, the dominant de ployment pattern for RAG remains a fixed single pass architecture in which a query is submitted once to a retrieval index, a static document set is returned, and an answer is generated from that set without any mechanism for self assessment, iterative refine ment, or adaptive re-retrieval. While this architecture is adequate for simple factoid questions whose complete answer can be recovered from a single retrieved document, it is fundamentally incapable of handling multi-hop questions, which are queries that require sequential reasoning across two or more documents where the answer to one intermediate reasoning step serves as the necessary context for formulating the next retrieval query. Failure mode is well characterised and systematic: If the single re trieved set is not complete (missing bridging facts), the generator either fabricates a plausible sounding but incorrect answer, or fails to complete an answer at all, without mechanisms to seek the missing evidence. This thesis directly tackles the above fault by designing, implementing, and eval uating thoroughly a Large Language Model (LLM) Driven Agentic Framework for Adaptive Retrieval and Multi-Hop Reasoning (MR). Three cooperative agentic mech anisms are introduced to the retrieval loop, adding to the passive fixed pipeline RAG architecture and making it an active, self-aware, evidence seeking process. The first is a question classifier that allows to classify an input query as simple or compositional and processes it through the right retrieval path, without causing the degradation of the accuracy that would result if complex decomposition strategies were applied to simple questions. The second is an iterative query rewriter that, at each retrieval step, explic itly identifies the specific information missing from the accumulated evidence set and constructs a targeted sub-query to retrieve it, appending only new documents to pre vent redundant context growth and terminating when no new evidence is returned. The third is a dual-signal sufficiency checker that combines a dense retrieval cosine simi larity threshold with an LLM based answerability judgement, requiring both signals to be satisfied simultaneously before answer generation is initiated, thereby reducing the probability of generating from insufficient evidence. The framework is evaluated on one hundred questions from the HotpotQA distrac tor benchmark, a dataset specifically designed to test multi-hop retrieval and reasoning under realistic retrieval noise conditions, achieving 46.0% Exact Match and 58.66% F1. These results represent absolute improvements of twenty four and sixteen per centage points respectively over a Vanilla RAG baseline, corresponding to a relative iv improvement of 109% in Exact Match, and outperform all three fixed strategy base lines compared. A structured ablation study establishes that all three components are necessary and complementary, with query rewriting identified as the single most crit ical component whose removal causes a ten percentage point Exact Match drop and a twenty five percentage point Retrieval Recall collapse. A comprehensive failure analy sis demonstrates that the framework successfully shifts the dominant error mode from retrieval failure to reasoning failure, directly redirecting future research effort toward generation quality rather than retrieval design. Together, these findings establish the proposed framework as a principled and empirically validated solution to the multi-hop retrieval and reasoning challenge that is practically deployable without any annotated reasoning supervision or model fine-tuning.