To develop a tool that helps physiotherapists to obtain specific and up-to-date information. It should overcome the problems of general-purpose models: Information must be up-to-date, specialized, understandable, and avoid hallucinations. A local system should guarantee strict data privacy.
Based on an open-source language model and a solid medical knowledge base, a Retrieval Augmented Generation (RAG) system was developed. The language model (LLM) was linked to a high-speed database. This database contains texts from PDF documents processed using natural language processing (NLP). The PDF documents were found and downloaded using DuckDuckGo. A query interface for the PubMed database was also integrated. Experts in physiotherapy designed the queries and developed and analyzed various test scenarios to evaluate the system.
The system works like a chat that receives inquiries and answers them. It uses a medically-based language model and carries out specific queries via DuckDuckGo and PubMed. The relevant documents were downloaded and processed using NLP to make them accessible to the RAG system. The evaluation shows that the system can generate specific, up-to-date answers. Over 85% of the experts were satisfied with the performance.
The local operating system summarises current publications from the Internet and the PubMed database and provides simple, understandable answers. Further research is needed to systematically identify the domain-specific needs of users and improve the system.
Applying local large language models requires some technical resources, but these can be realized on laptops. A customized RAG system allows physiotherapists quick and easy access to up-to-date knowledge in their field.
AI
large language models
