<p>The impeccable performance revealed by extractive QA systems in predicting answers to questions from a document corpus is highly appreciated. Nevertheless, the answers returned by extractive QA systems often lack relevance to the specific context of the questions asked. This discrepancy primarily arises from the fact that answers appear in different contexts across the collection of documents. To address this challenge, we propose a Context-driven QA (ContextQA) system that leverages the pre-trained BERT-based model RoBERTa at various stages. Further, the model is fine-tuned on custom narrative text datasets for the intended QA task. Specifically, the proposed model employs a three-stage data processing pipeline. During the first stage, the question phrase is transformed into a topic sentence through a process known as query-to-topic conversion. Consequently, the documents relevant to a specific topic are retrieved from the corpus for context-span extraction. In the second stage, for each relevant document, topic segmentation is applied to delineate the boundaries of potential context spans. This is achieved by computing relevance scores between the topic sentence and each context span from the retrieved document. Finally, the top <i>k</i>-ranked context spans are selected for the task of answer-span prediction. To accomplish these objectives, we use BERT-based language models and other NLP techniques across various levels and stages. The effectiveness of the proposed ContextQA-BERT is critically evaluated using a corpus of Indian epic stories, including the Ramayana and the Mahabharata. The experimental results indicate that the three-stage ContextQA-BERT system achieves high accuracy in extracting answer spans from the provided context paragraphs.</p>