Retrieval-Augmented Generation (RAG) grounds large language models in external evidence, but many RAG systems represent knowledge either as flat text chunks or as automatically constructed indexing graphs. This assumption is incomplete for curated thematic corpora, including religious scriptures, legal codes, clinical guidelines, educational taxonomies, policy documents, and library classification systems, where domain experts have already organized knowledge into thematic paths and citeable canonical units. This paper investigates how RAG can exploit such expert-authored structures while pruning evidence to a compact and query-specific set. We conduct a critical survey supported by a bibliometric analysis of 2815 Scopus-indexed RAG-related records exported on 26 May 2026, of which 2809 records were retained after duplicate removal. The bibliometric results indicate rapid growth in RAG research but limited explicit consolidation around curated thematic paths, canonical evidence units, or thematic path-guided evidence pruning. We therefore propose ThemePath-RAG, a retrieval framework that retrieves curated thematic paths as high-recall semantic routes, expands candidate canonical evidence, and applies query-aware scoring and global pruning before generation. To assess operational feasibility, we implement ThemePath-RAG for Qur’anic question answering and compare it with a Vector RAG baseline on 150 paired questions using RAGAS context relevance with gpt-4o-mini as the LLM evaluator. Both methods return approximately three final ayat per question. Vector RAG achieves higher mean context relevance than ThemePath-RAG (0.920 versus 0.798; p<0.001). Thus, the proof of concept establishes the feasibility of thematic-path-guided retrieval and identifies evidence-selection challenges, rather than demonstrating superiority over conventional vector retrieval. The paper clarifies the framework’s relationship to GraphRAG, LightRAG, HippoRAG, PathRAG, ontology-based RAG, and AI-augmented bibliometric systems, and outlines a language-matched, multi-baseline evaluation agenda for future cross-domain validation.
Winda Monika, D. A. Dewi, A. Nasution et al.· Information· 1 citation
Large Language Models (LLMs) are increasingly used in settings where the original user prompt may be unavailable, redacted, or untrusted. This makes response-only adversarial detection an important but challenging problem, because the detector observes only the generated response while the label is determined by the originating prompt. In this study, we investigate response-only adversarial classification of LLM outputs through a comparative evaluation of lexical, contrastive, and contextual representations. Specifically, we compare TF-IDF features, triplet-loss-based metric embeddings, TF-IDF+triplet concatenation, and a Sentence-BERT contextual embedding baseline across three tasks: Prompt Type, Adversarial Type, and Adversarial Class. Experiments on 158 592 response records show that response-only detection is feasible but substantially more difficult than prompt-aware detection. Under stratified group 10-fold evaluation, the strongest TF-IDF+XGBoost baseline achieves 72.37% accuracy for Prompt Type, 82.06% for Adversarial Type, and 56.40% for Adversarial Class. Leave-one-LLM-out evaluation further shows that cross-model generalization remains challenging. Triplet-based metric learning provides model-dependent benefits, particularly for some Adversarial Type configurations, but it does not uniformly outperform strong lexical baselines. Low-dimensional embedding analysis shows that Insertion samples are relatively separable, whereas Prefix and Suffix samples substantially overlap. Overall, the findings provide a realistic benchmark for response-only adversarial detection and clarify the limits of contrastive representations when prompt information is absent.
A. Nasution, Ahmet Emre Ergün, Aytu˘g Onan et al.· IEEE Access· 0 citations
This survey provides a comprehensive analysis of adversarial prompting strategies, ranging from input manipulation techniques to semantic and structural distortions, and explores defense strategies across preprocessing, model-level, postprocessing, and hybrid strategies, highlighting recent advances and their limitations.
A. Nasution, Ahmet Emre Ergün, Aytu˘g Onan et al.· IEEE Access· 0 citations