Retrieval-augmented generation is a data-quality project nobody scoped.

The pitch for retrieval-augmented generation is that grounding a model in your documents makes it accurate. The unscoped clause is that the answer is only ever as good as the documents it retrieves.

What happened is that an architecture got mistaken for a guarantee. The original method, introduced by Lewis and colleagues in 2020, pairs a model with a retrieval step so answers draw on an external knowledge source rather than memory alone. Vendor guidance is candid about the dependency that follows. AWS states plainly that grounded responses are only as reliable as the documents, databases, or APIs they are based on.

It matters because retrieval quality is a data-quality question wearing an infrastructure costume. Microsoft's own guidance notes that RAG quality depends on how content is prepared for retrieval. So the moment a corpus contains three definitions of a term, two stale policies, and a deprecated number nobody flagged, the retriever does its job perfectly and surfaces the contradiction with a citation attached.

What it reveals about the field is the recurring substitution of plumbing for governance. A survey of RAG robustness catalogs the failure honestly: retrieved context can be noisy, incomplete, or adversarial, and the system inherits those flaws. None of that is fixed by a better vector index. It is fixed by deciding what the authoritative document is and who maintains it, which is ownership, not retrieval.

What to watch is who owns the corpus. If retrieval accuracy is filed as a model problem, the team tuning embeddings will be asked to compensate for definitions nobody assigned. The reveal: a model grounded in undefined data is not careful. It is confidently wrong, and now it shows its work.