An answer should show its sources

Retrieval is not valuable because it decorates an answer with links. It is valuable when the system can show which claim came from which source, and when a person can inspect enough context to decide whether the answer is justified. Citations are part of the interface, not a footnote glued on after the model has spoken.

My course-material assistant work followed that shape. PDFs were processed with OCR and structure-aware retrieval, then answers carried page references, figures, confidence states, and missing-evidence signals. The product question was not only whether the model could answer. It was whether the user could check the ground under the answer.

A serious retrieval agent needs a calm way to say that the corpus does not support an answer. That state should be ordinary, not dramatic. The system can show what it found, what it can answer, and what source would be needed before it can go further.

This matters for documents, courses, contracts, reports, and internal knowledge bases. A fluent unsupported answer is worse than a useful gap report. When missing evidence is explicit, the agent can help the person fix the corpus or narrow the question instead of inventing confidence.

Answer quality and citation quality are different measurements. The answer can be right with weak support, or polished while citing the wrong paragraph. Evaluation sets should include direct answers, partial answers, duplicate documents, outdated material, and questions with no answer in the corpus.

The deeper pattern is the same as the lab's agent work: carry evidence through the system. Claims, sources, permissions, and review decisions should not dissolve into a transcript. They should survive long enough for someone to inspect, correct, and improve the machine that produced them.