Build multilingual RAG in my app with local neural retrieval and MandAPI only for answer generation. Use SentenceTransformer sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 and CrossEncoder cross-encoder/mmarco-mMiniLMv2-L12-H384-v1, or my explicitly selected compatible local model paths. Preserve the app and keep the real key server-side. Implement a --demo mode with fixed stubs and no downloads, heavy imports or API calls. The --live path must use normalized neural embeddings, top-k retrieval, CrossEncoder reranking and one POST to https://api.mandapi.com/v1/chat/completions with an exact account-visible MODEL_ID. Do not substitute TF-IDF and call it neural retrieval, and do not invent MandAPI embeddings/rerank endpoints. Check input token limits, keep stable source IDs and ask the model to answer only from retrieved passages with citations. Prepare installation and weight-download steps for me to run; do not download large weights or execute paid inference automatically. Report offline validation separately from neural and live results.