Most assistants answer the way an overconfident colleague does — fluently, and without telling you where the claim came from. That’s fine until the claim is wrong, and by then the fluent delivery has already done the damage.
Citations by default flip the burden. Instead of asking users to trust an answer, we hand them the evidence and let them verify it. Every synthesized response points back at the exact passages behind it, each with a source and a confidence score.
What a cited answer looks like
The shape is simple: a claim, then the receipts. Nothing is asserted that can’t be traced.
| Field | What it carries |
|---|---|
source_url |
Where the passage lives |
source_title |
Human-readable name of the source |
score |
How strongly it supports the claim |
When the evidence is too thin to clear the bar, the honest move is to say so rather than paper over the gap. That refusal is a feature, not a failure — it’s the difference between a system you can trust and one you merely hope is right.