https://www.mdu.se/

Long, multimodal technical documents are becoming more difficult to search reliably because evidence can be scattered across sections, figures, and tables. When conventional Retrieval-Augmented Generation (RAG) breaks documents into smaller chunks, the document structure is heavily flattened and structural relationships are lost. As a result, a typical Large Language Model (LLM) tends to hallucinate more during answer generation when it comes to processing large, complex documents. This thesis addresses this problem and proposes a Stateful Agentic Framework for Retrieval-Augmented Generation (SAF-RAG) to preserve structural metadata and recover missing evidence through bounded agentic control. A five-condition ablation ladder was evaluated on two representative datasets, LongDocURL and Visdom, under text-only and multimodal ingestion. It progressively introduced structural expansion, reranking, and a planner-navigator-reviewer agentic loop on top of state-of-the-art retrieval techniques. The main goal of this thesis was to improve answer groundedness and faithfulness while striking the right balance between reliability and cost. Each of five conditions was therefore evaluated for retrieval quality, lexical answer overlap, judge-based quality, and processing cost. Structural reranking consistently produced the biggest gains in retrieval quality. Across both datasets and ingestion variants, Normalized Discounted Cumulative Gain (nDCG) gains ranged between 11.3% and 44.4%, while Token F1 improved by 6.4% to 35.6%. Bounded agentic review then increased the combined LLM-judge score by 11.8% to 18.9% over structural reranking, although multimodal LongDocURL faithfulness declined by 13.6%. The results support structural reranking as the most stable choice for quality and cost and reserve agency review for cases where further judged answers justify mixed faithfulness and added computation.