RAG systems need better questions before better answers

A retrieval-augmented system can sound fluent while using the wrong evidence. That is why I treat retrieval as an inspectable information system rather than a hidden step before generation.

The first design question is not which embedding model to use. It is what a reliable answer should be able to point back to. Documents need stable identity, useful structure, and metadata that reflects how people actually search. Chunking should preserve meaning instead of cutting every file into equal pieces.

Evaluation also needs real questions. A small set of representative queries with expected sources is more useful than a large synthetic score that nobody can interpret. I check whether the right document was retrieved, whether the relevant passage appeared early enough, and whether the final response stayed within the evidence.

When evidence is missing, the system should say so clearly. A useful RAG application is not the one that answers everything. It is the one that makes the boundary between known, retrieved, and uncertain information easy to see.