ASTRUCTURAL AND STATEFUL APPROACH TO RETRIEVAL AUGMENTED GENERATION FOR CONTEXTUAL REASONING INKNOWLEDGE-INTENSIVE SYSTEMS

Abstract

Systems that are built around large language models(LLMs) have to answer questions whose answers are provided by the model itself. Its accuracy degrades on hierarchical collections, which are revised over time, and that demand exact citations. This thesis develops a framework for such collections in two parts. The first part is a structural pipeline that respects the hierarchy of the source documents at every stage. It splits text into sections and clause boundaries, attaches citation-grade metadata to each chunk, runs a dense neural retriever and a sparse term-matching retriever in parallel, fuses their ranked lists, and reranks with A cross-encoder that uses the metadata as an extra signal, and constrains the language model to a fixed output format. The second part is a stateful agentic layer that wraps a lightweight controller around the structural pipeline. The controller maintains a working memory across retrieval calls, reflecting on the evidence collected so far, decides whether the evidence is sufficient, and issues a refined query when it is not. The structural pipeline is implemented and evaluated on a corpus of central Indian statutes against a benchmark of two hundred test queries whose ground-truth applicable provisions were prepared by a practising domain expert. The pipeline identifies the primary applicable citation in 94.0% of test queries (188 of 200; 95% Wilson interval [89.8%, 96.5%]) and returns the full applicable citation set in 82.0% (164 of 200; 95% Wilson interval [76.1%, 86.7%]). Every neural component is used without domain-specific fine-tuning, so the reported numbers reflect the pipeline’s behaviour in the absence of any domain-adaptation effect. The stateful agentic controller is presented as a fully specified design, composed from methods already published in the agentic and self-reflective retrieval literature; its end-to-end empirical evaluation is left for subsequent work.