Embedding Model Selection for Domain-Specific Retrieval-Augmented Generation: A Comparative Study on Indian Cultural Heritage Corpora

This paper presents Kashivani, a domain-specific Retrieval-Augmented Generation (RAG) framework for Indian cultural heritage knowledge retrieval focused on Varanasi, India. The study evaluates three sentence transformer architectures, namely: all-MiniLM-L6-v2, all-mpnet-base-v2, and paraphrase-multilingual-MiniLM-L12-v2 across different corpus scales using manually validated question-answer evaluation. Experiments conducted on a corpus of 15 cultural heritage documents containing 2,876 chunks demonstrate that monolingual sentence transformer models consistently outperform multilingual alternatives for English-language Indian cultural heritage retrieval. The study further shows that embedding model rankings can reverse under corpus scaling conditions, highlighting the importance of evaluating retrieval systems at deployment-scale corpus sizes. The complete implementation includes document ingestion, chunking, FAISS vector indexing, semantic retrieval, and grounded generation using LLaMA 3.1 8B.