Retrieval-Augmented Pipeline for Clinical Pharmacology: From MeSH-Based Queries to Local LLM Answers

This thesis investigates the extent to which literature search and summarisation in clinical pharmacology can be partially automated by utilising a locally deployed Large Language Model (LLM) coupled with a Retrieval-Augmented Generation (RAG) component, designed as a free-text clinical decision support tool for medi- cal experts. The thesis aims to develop a prototype system consisting of a locally deployed LLM coupled with an API to PubMed. The prototype system success- fully converted free text clinical questions into structured search queries, retrieving relevant articles, to support clinical decision-making. The workflow analysis en- abled the development of partly automated search strategies that reflect the clinical workflow. The locally deployed RAG developed by Morphik, coupled with ColPali embeddings improved multimodal processing, enhancing retrieval quality. Clear guidelines must be established regarding when and how it should be disclosed that an AI system contributed to a patient’s treatment plan, either for individual pa- tients or for broader clinical recommendations. This work addresses a real and pressing need to enhance workflows with AI-assisted tools, without compromising on quality or accountability.