SemLab and Catharina Hospital develop RAG-tool for medical protocols

According to internal research, nurses at Catharina Hospital spend around 3 hours per department per day consulting medical documentation such as protocols and guidelines. While some of this time is spent reading and processing the relevant information, valuable time is lost searching through various sources and collecting fragmented data, instead of using it on patient care. Also, because nurses are under immense time pressure, scattered documentation might lead to it not being thoroughly consulted, or at all, which could degrade the quality and consistency of care. WIthin the context of the MedGPT-project, SemLab and Catharina Hospital are working together to combat this issue and develop a tool to decrease the time nurses spent on information gathering.

Large Language Models

In recent years Large Language Models (LLM’s) have become increasingly popular aides for (among other uses) question answering and information retrieval. More and more people rely on them for both personal and professional tasks. However, while undeniably powerful, out-of-the-box LLM’s are often designed with general purpose in mind and lack domain- and task-specific capabilities. For medical professionals at Catharina Hospital it is imperative only selected, trusted sources are used when looking up information. Which sources are used also needs to be transparent so answers are verifiable and relevant documents is easily accessible. Lastly, a very high degree of certainty that answers are correct is required as any wrong or missing information could have dire consequences. So models need to be overly cautious in whether they can safely answer a query, which contrasts with the directive of most LLM’s that aim to always have an answer for a user.

Retrieval Augmented Generation

To create a tool that fullfills these requirements, SemLab is developing a Retrieval Augmented Generation (RAG)-system specifically tailored for Catharina Hospital. A RAG-system is still a system that provides an answer to a question, but does so via a multi-step process. First, all relevant sources, provided by Catharina Hospital, are split into coherent chunks and stored in a vector database. When a user enters a query, a retriever-model gets chunks from this database that contain relevant information for this query. These chunks are then provided to a LLM as context based on which it needs to answer the query. Additional guardrails will be implemented around all steps to check that there are relevant sources available, the LLM indeed did base its answer solely on these, and there is a high degree of reliability and certainty for the answer.

Development

Initial development by SemLab is based on publicly available protocols from https://www.farmacotherapeutischkompas.nl/ and synthetically generated queries. The aim is to soon deploy a first prototype on-premise at Catharina Hospital. The goal of this rapid prototyping is for medical professionals to interact, test and experiment with the tool as early as possible, and on actual data from Catharina Hospital, as their feedback is essential to steer improvements and next steps. This development cycle combines the technical proficiency from SemLab and medical expertise and day-to-day use from Catharina Hospital to build a reliable and efficient tool for querying medical protocols.

Posted by SemLab on 17 april 2026