Clinical RAG with Docling and Presidio to speed up diagnosis
A diagnosis-support assistant that answers clinical questions from the institution's own documents (reports, protocols and internal literature) with cited sources, and with patient data anonymized before any processing.
The challenge
The clinical team lost time searching for information across scattered PDFs, reports and protocols, and no AI solution could be used without guaranteeing that patient data would neither leave the institution nor reach external models.
The objective
Cut the time between a clinical question and a grounded answer, without exposing personal data and without replacing the medical decision.
The solution
RAG pipeline: Docling converts PDFs, reports and tables into structured text; Microsoft Presidio detects and anonymizes personal data (names, IDs, dates) before indexing; hybrid search (vector + BM25) retrieves passages and the model answers with citations. Guardrails, evaluation sets reviewed by professionals and audit logging. The decision is always human.
The impact
- Answers with cited sources from the institution's own documents
- Patient data anonymized with Presidio before leaving the controlled environment (LGPD)
- Evaluation set reviewed by healthcare professionals to measure faithfulness
- Support tool: the clinical decision stays with the professional
Jean C Becker
Senior Solutions Architect | AI & Machine Learning Specialist
Designed and built this work. About JBKR →