EcoTalkBot

A RAG chatbot for science communication around biodiversity

EcoTalkBot is a chatbot that was developed as part of a research project at Aarhus University, which aims to promote interest in and understanding of biodiversity on agricultural land. Through an interactive dialogue with the chatbot, users have easy access to the latest knowledge about biodiversity. The answers are based on a selection of reliable sources, made available through a RAG system. Participants had a conversation with the chatbot and filled out a survey. The researchers learn about what questions people have and how a conversation with a chatbot can affect their perceptions.

Features:

• Multi-lingual: While the interface is in Danish the project is set up to be multi-lingual. The source documents are in a number of different languages, and the bot is instructed to respond in the language the user writes in. Having source documents in different languages, means that it is important to use an embedding model that is optimized for multi-lingual search.
• Deployed in Docker on Azure, using the Azure version of GPT-4o, and a Weaviate vector-database for search.
• User-friendly design with Streamlit: Sidebar with information about the project and bot; List of source all documents can be viewed in a tab; Example questions for getting started, one general and one more specific. Especially the more general suggested opening question was clicked by multiple participants.
• Verifiable references: The sources used in the response are listed in an expandable box, with links to the original documents. This list of references is produced by deterministic code, to make sure the details are correct.
• Login and data collection: The interactions were recorded, including user ID, user input, chatbot response, references, timestamp.