International Journal of Science, Engineering and TechnologyAuthors: Doke Sakshi Vitthal, Professor Bangar Abhishek P., Professor Karad Vandana A.
Abstract: AI chatbots built on large language models (LLMs) answer from parametric knowledge that is fixed when training ends, so they cannot reliably reflect updated, institution-specific or access-restricted information unless the model is retrained or connected to external sources. This paper presents a structured narrative review of 17 sources published between 2019 and 2026, centred on five peer-reviewed IEEE Access studies from 2024–2026 and 12 foundational works, organised into seven themes: the static knowledge bottleneck and knowledge injection; retrieval foundations and pipeline design; fine-tuning, retrieval-augmented generation (RAG) and hybrid comparisons; retrieval-bypassing and question-indexed alternatives; corrective and adaptive retrieval; permission-aware retrieval; and local deployment, evaluation and reliability. Reported evidence shows that retrieval substantially improves on closed-book generation, that fine-tuning and RAG have complementary cost and quality profiles, that a combined approach attained the highest semantic scores in the one study that tested it, that similarity-based answer validation and question-indexed retrieval improve grounding on domain corpora, and that per-user access control can be enforced at retrieval time with modest overhead. However, none of the five core studies evaluates a mechanism that routes queries between a parametric path and a retrieval path, none measures answer quality after the underlying knowledge changes, and most validation is limited to one corpus or benchmark. The paper contributes a thematic synthesis, a comparison of twelve representative studies, a proposed conceptual seven-layer dual-path architecture and a prioritised research agenda.
