RAG-Based Domain-Specific Search: Leveraging Reinforcement Learning in Large Language Models

ABSTRACT

As domain-specific information systems become more complex, traditional search and retrieval methods are starting to fall short. Most of these systems depend on static indexing, simple ranking rules, and methods that don’t adjust over time. As a result, they always struggle with accuracy, fail to adjust to new information, and does not have a strong understanding of context, which is often found as problems that are especially serious in high-stakes areas like healthcare, law, and finance.

This study tackles these issues by introducing a Reinforcement Learning–Enhanced Retrieval-Augmented Generation (RL-RAG) framework. The approach improves how documents are retrieved and how responses are generated by using a reward-based system that continuously learns and optimizes performance.

This study used an experimental approach based on a domain-specific dataset of 16,412 question and answer survey, drawn from different medical fields. The data combined ROUGE-L, Exact Match (EM), and token-level F1 scores.

Keywords: Reinforcement, Retrieval-Augmented Generation, Domain-Specific, Reward-Based, Adaptability.