A RAG team wants direct control over retrieval, reranking, prompt construction, and answer generation rather than one combined call. Which API pattern best fits?

1. Use only a combined generation call and assume every intermediate step can be replaced afterward.
2. Sort passages only by their file extension.
3. Skip retrieval and ask the model to reproduce the company private database.
4. Retrieve evidence separately, apply the approved ranking and context logic, then invoke the generator.