A Markov-Decision Formalization of Context Selection and a Hybrid Text–Graph Knowledge Model for Retrieval-Augmented Legal Question Answering

Retrieval-Augmented Generation (RAG) couples a large language model with an external knowledge store, yet the quality of its answers is dominated by an upstream, under-formalized step: which context to retrieve. This paper addresses two gaps. First, we give a precise formalization of relevant-context selection as a constrained information-maximization problem and reduce it to a Markov Decision Process (MDP), passing through a non-observable context-utility functional, a computable surrogate that combines relevance, diversity and graph coherence, and an analysis of its diminishing-returns structure that justifies both greedy and learned policies. Second, we propose a hybrid text–graph model of documented information that unifies a lexical, a vector and a graph level over a heterogeneous text-attributed graph, together with the database-structuring principles that make the three levels jointly searchable. We instantiate the model on the national legislation database of the Republic of Uzbekistan, building a typed three-level graph (documents, regulated entities, defined concepts) of 207 nodes and 1,718 typed edges directly from act cross-references and definition articles. On four multi-hop question-answering benchmarks and a 1,197-document legislative corpus, the resulting method (GraphRAG-RL) reaches an F1 of up to 0.832 and a domain accuracy of 91.3%, improving over naive, graph and reinforcement-learning RAG baselines at a moderate retrieval cost.

Hali tarjima qilinmagan