Listen: I think Google got it wrong with “Generate → Ground” approach.

When it comes to building large language models, should we write first and check our facts later, or find the facts first and write second?

Google developed a system called Retrofit Attribution using Research and Revision, or RARR. It represents the write-first approach. The model drafts an answer, generates search queries to fact-check itself, and then edits the text. While this is a clever way to patch up an existing model, it is incredibly fragile. If the search queries are off-target, the whole system fails. Plus, it is slow, and it often fixes errors simply by deleting useful information.

A much stronger alternative is Retrieval-Augmented Generation, known as RAG. RAG flips the order. It finds the evidence first, puts it directly into the context window, and then lets the model generate the answer.

This retrieval-first approach has major advantages. It has built-in factuality because the model reasons over real, retrieved text instead of guessing. It is faster because it only requires a single pass. And if the retriever finds nothing, the system can gracefully say "I don't know" rather than making things up. Advanced methods like Fusion-in-Decoder push this even further, letting models scale to larger sets of evidence without slowdowns.

If you are architecting a system from scratch, don't spend your time polishing a post-hoc band-aid. Grounding should always come before generation. Put the evidence in the context window first, and let the model write from a position of knowledge.