Every platform now offers "chat with your documents". Very few organizations still use it six months later. The gap is not the model — it is retrieval engineering.
What retrieval engineering covers
Chunking strategy that respects document structure. Metadata that enables permission-aware retrieval. Freshness: pipelines that keep the index aligned with reality. Contradiction handling: what happens when two documents disagree. Citations: so a user can verify rather than trust.
The honesty requirement
A knowledge system also needs the ability to say "I don't know." Systems that always answer are not knowledgeable — they are confident, which is worse. Escalation to a person with context attached is a feature, not an admission.
Evaluate retrieval separately
Model quality and retrieval quality are distinct variables. Evaluate them separately, fix them separately, and monitor them separately. Most knowledge-system failures are retrieval failures wearing a model costume.