Enhancing Hindi Question Answering in RAG Systems via Fine-Tuned Retrieval and Generation Models

Retrieval-Augmented Generation (RAG) systems enhance natural language understanding by retrieving relevant context before generating responses. However, due to poor embedding representations and generation errors, existing models struggle with handling Hindi questions. This work presents a fine-tuned Hindi RAG system using paraphrase- multilingual-mpnet-base-v2 as the retriever and facebook/mbart- large-50 as the generator. The retriever was trained with Multiple Negatives Ranking Loss to improve query-context alignment, while the generator was fine-tuned for concise, high- quality responses. Our model achieved an average retrieval similarity of 0.8700, ensuring effective query-context alignment, a BLEU score of 0.6833, reflecting high response quality, and a semantic similarity of 0.9200, demonstrating strong meaning retention. As expected, the exact match rate was 0, as the model generates fluent yet non-identical responses. These results confirm the system’s effectiveness in improving retrieval accuracy and response fluency for Hindi QA applications.