Self-Reflection and RAG in Small Language Modelsfor Enterprise Knowledge, Systematic Mapping Study

Abstract

Organizations require artificial intelligence systems capable of reasoning over internal knowledge bases without compromising security. However, large language models (LLMs) do not natively access confidential information, and their local deployment is often impractical due to hardware costs. In this context, small language models (SLMs, 1B–13B parameters) emerge as a viable alternative, although their ability to support retrieval-augmented generation (RAG) pipelines with self-reflection mechanisms in enterprise environments remains an open question.

This study analyzes the integration of RAG and self-reflection in SLMs through a systematic mapping study (SMS). Following the guidelines of Kitchenham and Charters (2007) and Petersen et al. (2008), publications from arXiv, NeurIPS, IEEE Xplore, and ACM between 2020 and 2025 were examined. From an initial set of 510 results, 41 primary studies were selected.

The results reveal a growing interest in architectures that combine dense retrieval with adaptive strategies based on uncertainty, where a high of the approaches rely on integrating existing components. Furthermore, limitations are identified in evaluation, efficiency, and applicability in real-world enterprise scenarios. It is concluded that these techniques show potential to improve reasoning and adaptability in language models; however, significant challenges remain for their implementation in SLMs under privacy and hardware constraints. This study organizes existing knowledge and establishes a structured foundation for future research.