Bridging the Gap in Mental HealthLiteracy: Evaluating RAG Architectures for Scientific Research Communication

Sammanfattning

As mental health disorders rise among youth, improving the general public’s ability to navigate health information is essential for ensuring timelysupport.Thisthesis investigated whether Retrieval-Augmented generation and Chain-of-Thoughtprompting can serve as effective methods for developing a conversational system aimed at improving mental health literacy. A multi-agent pipeline was implemented, comprising a retrieval agent and an answer-generation agent. Evaluation was conducted using a combination of automated metrics, including faithfulness, answer relevancy, and readability, as well as human evaluation. The resultssuggest that retrieval-augmented generation can effectively ground answers indomain-specific knowledge, achieving high faithfulness and minimizing hallucinations. Chain-of-thought prompting improved the structure and transparency ofresponses. The findings also demonstrated that language models can adapt explanations to non-expert users; however, improvingreadability introduced a trade-offwith faithfulness. In conclusion, the combination of retrieval-based grounding andstructured prompting represents a promising approach for communicating scientific information to non-expert audiences. However, balancing accuracy, readability, and educational value remains a key challenge for future systems.