In today’s business, knowledge is one of the most valuable assets – especially for organisations like the Ausgleichskasse Basel-Stadt (AKBS), which deals with complex questions on a daily basis. In the family allowances department, however, this knowledge was spread across the internal wiki, intranet documents (including guidelines and directives) and the online help within the specialist application.
In order to consolidate and leverage this knowledge, the AKBS relied on Netrics' AI Jumpstart programme. In a kick-off workshop, ten questions were defined to test whether a “ChatGPT-like” solution could reliably answer domain-specific questions.
The proof of concept was carried out in two steps: First, all core documents were fragmented, converted into vectors and stored in a vector database. This allowed user queries – which were also transformed into vectors – to be matched with relevant text passages through Azure Cognitive Search and subsequently converted into answers using Azure OpenAI.
The initial version of the solution was functional but not yet precise enough. In the second step, the documents were therefore refragmented and divided into smaller vectors. The team also switched to a fully vector-based search and refined the prompt engineering. The result was a significant improvement in answer quality – ten out of eleven test questions were answered correctly.
The chatbot now enables employees to retrieve information much faster, increasing productivity across the department. New employees also find their feet more quickly thanks to the AI solution. Based on the successful proof of concept, AKBS has decided to implement further AI use cases together with Netrics.