Citations Get Checked, Not Trusted

A citation is only as good as the validation behind it. It's easy to have a model emit something that looks like a citation — a bracketed number, a document name — without any guarantee that it actually points at retrieved content. GroundedDocs treats that gap as the primary failure mode to design against.

Structured answer-with-citations as the default, not an add-on

The model returns structured JSON with citation IDs bound to specific retrieved chunks. This isn't a post-processing step layered on top of a free-text answer — it's the shape of the output from the start. Every citation ID has to resolve against the chunk set that was actually retrieved for that query.

Validation before an answer ships

A validation pass runs before any answer is allowed to reach the user: it checks that every citation ID in the response resolves to a real, retrieved chunk. If a citation doesn't resolve, the answer is rejected. An unvalidated or broken citation is treated as a defect in the system, not a cosmetic detail to smooth over in the rendering layer.

Holding the line with a frozen eval set

To make sure this behavior doesn't quietly drift as prompts or the underlying corpus change, the system is held against a frozen 40-question set: 18 answerable, 11 partial, 11 unanswerable. Because the set doesn't change, a regression in citation faithfulness shows up immediately in the eval gate rather than surfacing weeks later as a user complaint. Citation faithfulness is tracked as its own metric, separate from retrieval quality — a system can retrieve the right chunk and still fail to cite it correctly, and that distinction matters for debugging.