A patient portal and a research layer that cites its sources
Fertility care runs on evidence. We built the retrieval and the interface that put a decade of it within reach of clinicians and patients.
Objective
Nova's clinical team fielded the same questions repeatedly, each answer requiring someone to find the right passage in a decade of literature and protocol documents.
They were explicit that a confident-sounding summary would be worse than nothing: any answer had to point at its source and refuse when the corpus did not support it.
Solution
We audited the corpus first — formats, duplication, contradictions and what was simply out of date — and agreed a source of truth where documents disagreed.
Retrieval was designed against document structure, combined with keyword and metadata filters, and measured separately from the model's writing.
The eval set was written with Nova's researchers and handed to them, so quality is their number rather than ours.
Outcome
Clinicians ask in plain language and get an answer with the passage attached, or an honest refusal. The patient portal reuses the same layer under clinical review.
Access policies are enforced at query time, so retrieval cannot surface material a role should not see.
What’s next
Extending the same evaluation discipline to the intake questionnaire, and a clinician feedback loop that files low-confidence questions for review.
“It tells us where the answer came from, and it tells us when it doesn't know. That was the requirement.”Nova FertilityClinical research lead
