Case Study
Medical Chatbot to Support Patient Research
With a chatbot based on self-defined, scientifically validated content, patients can be provided with accurate information for their own research.
Background
A government-licensed medical service provider was confronted with a high volume of recurring, often simple patient enquiries. These reached the organisation via email, telephone, and in-person contact, tying up qualified staff and leading to delays in processes as well as rising operational costs. Following our consultation, the client decided to implement a proof of concept (PoC) for a RAG-based AI assistant to support patients in conducting independent research and to provide professionally accurate, source-backed responses.
Requirement
The goal was to create a conversational assistant integrated into the customer portal that would generate responses solely based on a curated knowledge base (medical pamphlets, information leaflets, relevant professional literature) provided by the client, while transparently citing the sources used. Where appropriate, ICD-10 references were to facilitate further research. The knowledge base needed to be expandable by the client’s own medical staff without the involvement of js-soft (via document upload and web content ingestion). At the same time, it had to be ensured that no personal data would be processed; optionally, care staff bound by medical confidentiality could, with the explicit consent of the affected person, access their queries in order to prepare for consultations.
Solution
The js-soft project team implemented a multilingual RAG architecture and integrated it as a dedicated section within the customer portal. Documents (PDF, DOCX, HTML) and specified web pages are processed via an ingestion pipeline with text extraction, semantic chunking, and metadata enrichment, then indexed in a vector database.
The search function combines semantic similarity with optional lexical weighting and reranking to ensure both relevance and coverage. A large multilingual LLM generates responses strictly within context and includes clickable section-level citations.
Medical terms are mapped to ICD-10 codes using NER and terminology matching; in cases of unambiguous assignment, both short description and code are shown; in uncertain cases, a neutral reference to relevant chapters is provided. Guardrails prevent the issuance of diagnoses or therapy recommendations and make it clear that the chatbot serves as a tool for research and orientation. Operationally, multi-tenancy, role-based access, end-to-end encryption, and audit logging are implemented. Upon request, patients receive access to a personalised pilot version; consent for staff to view patient questions is granular, revocable, and logged.
The PoC was evaluated using a jointly defined catalogue of reference questions and expert reviews and was approved following successful assessment.
Benefits for the Company
The assistant significantly reduces the effort required to respond to repetitive enquiries and speeds up communication with patients.
The solution can be rolled out quickly as a non-invasive, cost-efficient pilot.
The autonomous maintenance of the knowledge base by the medical team ensures long-term scalability, while consistently cited original sources and ICD-10 references where applicable enhance transparency and trust.
We listen, we understand, and we execute
Let’s discuss your requirements – together, we’ll find the best solution for your business. Looking forward to our conversation. Estelle Hounsa, Sales Managerin js-soft