A Retrieval-Augmented Generative Artificial Intelligence Framework for Domain-Specific Knowledge Discovery and Reliable Question Answering

Abstract

Domain-specific question answering requires both linguistic fluency and evidencegrounded accuracy, yet conventional large language models may produce unsupported claims, omit document provenance, or fail to reflect updated institutional knowledge. This paper proposes a retrieval-augmented generative artificial intelligence framework that integrates document normalization, metadata-aware semantic chunking, hybrid sparse–dense retrieval, cross-encoder reranking, evidence-constrained generation, citation verification, and selective abstention. A modular mathematical formulation defines retrieval fusion, context selection, grounded decoding, and confidence-based response control. The evaluation protocol specifies document-level partitioning, expertannotated questions, retrieval and answer metrics, latency measurement, and error analysis. To illustrate reporting requirements without fabricating empirical evidence, numerical tables and graphs use explicitly hypothetical example values; these are not measurements of an implemented system. The framework provides a reproducible architecture and a testable set of hypotheses concerning answer correctness, evidence faithfulness, retrieval coverage, and operational efficiency. The contribution is a transparent research design for trustworthy knowledge discovery in organizational, technical, academic, and regulatory collections.