New models and features improve Svea

More than 60 municipalities, regions, and public authorities are involved in the project and recently received an updated roadmap for the remainder of the year. The work is organized into four focus areas: training and change management, improving the user experience, advancing AI models, and resolving a number of legal questions.

A key insight from the project so far is that managers and leadership teams have an important role to play in creating incentives for all employees to start using generative AI. Leaders need to lead by example, set clear priorities, and allocate resources toward knowledge building and organizational change.

“Our most active Svea users are those who have received a proper introduction to generative AI. They understand how to use Svea, its limitations, and have a clear idea of purpose. That’s why we will focus more on supporting the leaders involved in the project, so that they, in turn, can foster the right conditions within their organizations,” says Jonatan Permert.

Joakim Uddling, digital strategist at Sollentuna Municipality and responsible for the municipality's participation in the project adds:

Annotation lays the foundation for better answers from Svea

This spring, project participants have collectively spent 2,500 hours annotating training data. In practice, this means rating combinations of questions and answers, data which is now used to train the AI models behind Svea.

The retrieval model is at the core of Svea’s 'knowledge source' feature. When activated, it enables the chatbot to base its answers on information from an available document collection. Depending on the user's organization, this could include municipal guidelines for elderly care, national legislation, administrative guides, and more. 

“This type of solution is known as RAG – retrieval-augmented generation, which combines two types of models: a retrieval model that fetches relevant facts from documents, and a language model that summarizes and formulates an answer,” says Adam Ek. 

In other words, the factual accuracy of Svea’s answers depends on the quality of the retrieval model: the better the retrieval model, the more accurate the answers. Jonatan Permert explains why the project team decided to develop their own: 

“The existing embedding models available for Swedish are simply not good enough for the types of documents found in the public sector. To create a tool that truly adds value in this area, we need to train a better model – and that’s also where the power of collaboration comes in. No single participant in this project could have done this alone—not in terms of creating the data or training the model. But together, it becomes possible.”

New features and legal resolutions

Among the upcoming features are voice capabilities – both for speaking prompts and transcribing audio files. Both have been highly requested by Svea’s current users. Another area of development involves 'specialists': dedicated chats for specific tasks, such as project documentation, HR inquiries, or communication support.

“To help users get started, we are also working on making the interface more intuitive and specifically adapted to the needs of the Swedish public sector. Since the vast majority are inexperienced or first-time users of generative AI, it is especially important that Svea feels accessible and easy to use,” says Jonatan Permert. 

The project also addresses a number of legal questions surrounding the use of generative AI in the public sector.

“Among other things, we are currently investigating whether it is legally permissible to process sensitive personal data and confidential information in Svea. In the long run, we’re hoping to be able to support use cases that involve this sort of data,” says Jonatan Permert.