Reading up on CrossEncoder

Production RAG systems fail with naive fixed-size chunking and top-k retrieval, so the author recommends semantic chunking, hybrid search with BM25, reranking, context expansion, and a rigorous evaluation framework to improve accuracy and reliability.