Although Large Language Models (LLMs) have progressed significantly, Retrieval-Augmented Generation (RAG) still falls short for under-resourced Indic languages. Existing embedding systems, built primarily for English, face issues with script mismatches and limited word coverage, leading to near-complete retrieval failures in mixed Hinglish contexts. This paper introduces Hybrid Indic-RAG, a simple framework to reduce meaning shifts. It uses a targeted word-matching layer to connect Romanised Hinglish with standard Devanagari Hindi, paired with a two-part scoring system that favours exact word matches over fuzzy semantic signals. Tests in fields like healthcare and books show it cuts retrieval errors by 45% and lifts fact accuracy to 75%, well ahead of standard RAG methods. Our approach offers a low-cost way to build stable, error-free question-answering tools for India's language ecosystem.
It is found that a reasoning model without any retrieval augmentation outperforms every RAG variant the authors tested, including catalog grounding via Lucene fuzzy matching, brand-model predictions, click/conversion signals, and web-search retrieval.
This work compares Turkish document question answering across three chunking strategies, five embedding models, and two LLMs, over three documents with contrasting layouts, finding the faster LLM is not the more accurate one.
Mustafa Sertac Turkel, Fatma Nur Korkmaz, Ahmet Tugrul Bayrak· 1 citation
A systematic analysis of existing evaluation frameworks and metrics for RAG-based Question Answering (QA) systems reveals urgent needs for hybrid evaluation frameworks, reasoning path traceability, combined pipeline assessment, and harmlessness evaluation, laying groundwork for the next generation of evaluation methodo...
A structure-aware RAG framework is proposed that models Bengali textbooks as hierarchical graphs and uses a contrastively trained graph neural network to retrieve a small set of relevant passages, enabling topic-specific multiple-choice question (MCQ) generation and in-domain answer prediction.
Abu Tarabin Surzo, A. K. M. Nihalul Kabir, Sm Azmain Faysal et al.· 0 citations
It is demonstrated that model selection, retrieval behaviour, and hardware limits must be evaluated separately when deploying open-weight LLMs for Turkish domain documents.
Imtiaz Ul Hassan, Öykü Akbulut, Onur Kaya et al.· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
