RAG without citations is just a guess with extra steps
Retrieval-augmented generation gets described as a way to give a model your data. It is more useful to think of it as a way to give the model a reason to be right, and reasons have provenance.
Retrieval-augmented generation gets described as a way to give a model your data. It is more useful to think of it as a way to give the model a reason to be right, and reasons have provenance.
Citations are not decoration
When an answer carries the source it came from, three things change. The reader can check it. The failure becomes visible rather than plausible. And the team gets a debugging trail: a wrong answer with a citation is a retrieval bug you can reproduce, while a wrong answer without one is a mystery.
If your interface has room for one feature beyond the answer itself, make it the source link.
Chunking is a product decision
The instinct is to treat chunk size as a tuning parameter. In practice it decides what a citation can point at.
Split by paragraph and a citation lands on a sentence with no context. Split by document and it lands on forty pages. Split at the section heading, and the reader arrives somewhere they can orient themselves — which is the whole point of the link.
What to do when retrieval finds nothing
Say so. The system that answers anyway, from general knowledge, is the one that eventually tells a customer something confidently wrong about their own contract. An assistant that occasionally says "I could not find this in your documents" earns more trust than one that never does, because the times it does answer mean something.