AI-Cares

In the VLAIO-TETRA project AI CARES, Artevelde University of Applied Sciences, Thomas More University of Applied Sciences and Ghent University are developing chatbot prototypes trained on specific knowledge sources from the fields of welfare and mental health care. Across several living labs, we test different techniques and, together with organisations from the field, explore how AI chatbots can support better conversations about psychosocial topics.

We experiment with techniques including system prompting, RAG, structured decoding and LoRA fine-tuning. With system prompting, we provide the chatbot with clear instructions in advance about its role, tone, boundaries and way of working. With RAG, the chatbot retrieves relevant information from selected knowledge sources so that its responses are better grounded in reliable content. With structured decoding, we define in advance the structure according to which the language model must generate information. With LoRA fine-tuning, we adapt the underlying language model using example data, without retraining the entire model.

Through the WatWat chatbot, developed together with De Ambrassade, we investigate how a chatbot can better tailor information to the question and situation of the person seeking information, without compromising content quality. The Triple chatbot demonstrates how different ethical frameworks and boundaries can influence the way a chatbot responds to sensitive psychosocial topics. The OHUP Training Chatbot, developed together with SAM vzw, allows users to practise online counselling conversations through role-play with an AI chatbot. The chatbot uses training materials as its knowledge base. During practice, users can request AI text suggestions and receive targeted feedback. They can also add their own knowledge files, adapt system prompts and create new practice scenarios. The JAC Intake Chatbot, developed together with JAC East Flanders, helps young people clarify their support needs before the start of a contact. Together with the young person, the chatbot gathers relevant information about the support request and its context and then structures this information for the staff member. This gives the professional a quicker understanding of the situation and leaves more time during the conversation for deeper exploration, connection and support. The CLB Finetuned Chatbot allows us to compare whether adapting the underlying language model itself — by fine-tuning it on 14,000 example conversations from CLB Online — also leads to better chat conversations.

Try out the different prototypes yourself and draw your own conclusions. We would be happy to hear what you think.