RAG-Based Question-Answering Systems for Closed-Domains: Development of a Prototype for the Pollution Domain

In recent years, question-answering (Q&A) systems have become essential tools for combating misinformation and providing specific information to answer users’ questions. In this paper, we introduce the design of a Q&A system based on a retrieval augmented generation approach. This approach combines the power of semantic retrieval models to find relevant information from large textual corpora and the ability of pre-trained language models to process natural language and generate new answers. The workflow connects retrieved documents from an updated database containing domain-specific information and provides them to the generative module to produce reliable answers for users based on the context. The design of the proposal is notable for its adaptability to different domains, as the system dataset can be iteratively created from updated content based on Wikipedia pages. To test the proposal, we developed a Q&A prototype focused on the pollution domain. Through an iterative approach, the prototype was developed over three iterations, successfully addressing and overcoming key challenges and making significant progress in operational efficiency and answer quality. Based on the preliminary results obtained, the performance of the RAG-based prototype is satisfactory. After evaluating a small subset of questions, only 10% of the answers were not very coherent with the context provided.