Retrieval-Augmented Generation for Urdu Islamic Jurisprudence: A Review of Design Choices and a Reference Architecture

Questions about Islamic law need exact answers with sources. A wrong citation is a serious failure. Retrieval-augmented generation (RAG) fits this need, but Urdu is a low-resource language and there is little guidance for building such a system. This paper reviews the design choices behind an Urdu Fatawa RAG system and proposes a reference architecture. We compare Urdu-capable language models, embedding models, tokenizer cost, vector databases, primary data stores, and backend frameworks against four requirements: faithfulness to sources, precision on close jurisprudential terms, Urdu fluency, and private self-hosting. We argue for chunk-level storage of question-answer pairs, hybrid dense and sparse retrieval fused by reciprocal rank fusion, a prompt with English instructions and an Urdu output example, and a three-container deployment that separates inference, vector search, and the API. Our choices rest on published benchmarks, not on a controlled comparison on our own corpus, and some remain hypotheses. We state which ones, and we give an evaluation plan.