rag-integration makes RAG stages swappable

This is the right abstraction for RAG debugging: most failures live in glue code, not the model.

• –The `docs__...__evaluated` chain turns ablation into a normal workflow, which is exactly what you want when a chunker or embedder swap tanks retrieval quality.
• –The repo already includes text and image pipelines, FAISS retrieval, BEIR/Flickr30k loaders, and metrics like precision, recall, NDCG, and MAP, so the concept is backed by real scaffolding.
• –The big risk is semantic drift: a swap can be “typed” and still break downstream assumptions unless schemas, defaults, and versioning are enforced hard.
• –For production, I’d want explicit provenance and cache invalidation rules; otherwise this can still become a hard-to-debug pile of compatible-looking plugins.