Religious Debate AI solves RAG entity gaps

Building RAG for historical texts proves that pure semantic search fails at entity alignment and collection-size bias. Hybrid retrieval with keyword boosting is required to bridge name variations like Moses and Musa, while custom chunking strategies are essential to preserve complex citation structures like surah/ayah and book/chapter/verse. Source diversity tuning prevents massive collections from overpowering smaller scriptures, and constraining Llama 3.3 to citation-only answers effectively eliminates hallucinations. The project serves as a blueprint for handling heterogeneous datasets with strong internal referencing.