LLMs are used for everyday QA, but they can produce outputs that can be hallucinatory, making them inaccurate, especially when unverified results are used. A major limitation in Sanskrit literature, like the Mahabharat, is that it is difficult for generic models to grasp the metaphoric meaning of the text. Factual accuracy depends heavily on domain-specific knowledge needed for supporting standard frameworks, which are missing in today’s LLMs or RAG systems. To solve these problems, we introduce a Hybrid RAG method, which is trained to answer Sanskrit epic questions specifically. We use expertise in knowledge base retrieval and semantic embeddings, merging query categorisation and hybrid retrieval methods. The response is generated using the Gemini family of models for its core generation and evaluation tasks. The input is classified into one of the several RAG strategies. An impartial Gemini Pro model is used for judging; it evaluates the generated responses. Research on this strategy-driven, robust pipeline significantly increases factual accuracy and contextual relevance.