Förenkling av Offentlig Upphandlingsanalys med Stora Språkmodeller : Extraktion och visualisering av utvärderingsmodeller i offentliga upphandlingar med hjälp av stora språkmodeller

This thesis investigates whether it is possible to implement an algorithm that uses Large Language Models to extract complex evaluation components from Swedish public procurement documents. Current work involves manual analysis and manual calculation from experts in the field, taking up time and resources, and making it exclusive to domain experts. This solution simplifies the process and allows anyone to evaluate their bids on complex evaluation models in seconds. The problem was solved by implementing a multistage algorithm integrated with Google’s Gemini models that accurately and consistently extracts, displays, and allows users to simulate their bid’s value. The results demonstrated high accuracy for both simple and intermediate documents, with 90% correct extractions, while the most complex documents reached an accuracy of 50%. In conclusion, this approach clearly outperforms the current manual process in terms of both time and resource efficiency, and with further development, it has the potential to effectively handle even the more complex documents.