A question-answering layer over unstructured clinical documentation, built so that every answer cites the source passage a reviewer can open.
Clinical staff were searching across scanned referrals, discharge summaries and structured records held in three different systems. A first attempt at a chatbot produced fluent answers that staff could not check, so it went unused.
The requirement was not better prose but traceability. Patient data could not leave the customer's cloud tenancy, source documents were a mix of clean text and uneven OCR output, and the team had no ML engineers to maintain a bespoke pipeline.
Rebuilt ingestion so every chunk retained a stable pointer back to its document, page and character range, rather than asking the model to cite.
Clinical codes and drug names were matched far more reliably by keyword search than by embeddings alone. Re-ranking was added only where evaluation showed it changed the top result.
Three hundred real questions supplied and graded by clinical staff, scored on every pull request, so prompt and retrieval changes were measured rather than argued about.
A schema that required at least one citation, returning 'not found in the record' rather than an unsourced answer. The in-house team now owns the evaluation set and the deployment.
Clinicians use the assistant daily because they can check it in one click. The hard problem was never the model; it was making the provenance of an answer a structural guarantee of the pipeline instead of an instruction in a prompt. Once citation could not be skipped, trust followed.
Clinicians use it because every answer links to the exact passage it came from. That one decision made the difference between a demo and a tool.
Next story · B2B SaaS platform, Australia · SaaSRebuilding tenancy and billing without a migration weekend.Read next →Tell us about your system and we’ll give you an honest read on where to start.
Book an introductory call