A Bangla-language retrieval-augmented conversational agent for maternal health in a low-resource setting: design and formative evaluation

Abstract

Abstract Maternal health information in Bangladesh is difficult to access for women with limited literacy and no English, and perinatal mental health remains under-supported. We designed and formatively evaluated Baby and Me, a Bangla-language conversational agent that retrieves from a curated, physician-reviewed corpus of 247 maternal health articles, segmented into 2,945 indexed passages, and generates responses grounded in the retrieved material. A survey of 72 pregnant and postpartum women established design requirements; Bangla-only language preference was universal among women with no formal or primary education, and a quarter of households reported annual income below one lakh Bangladeshi taka. Candidate model configurations were screened against physician-authored reference answers before deployment. Twenty-four respondents then rated the deployed system: ease of use was highest (mean 4.79 of 5) and clarity strong (4.54), while trust (4.29), concern alignment (4.04) and empathy (4.00) were lower and more dispersed. Interviews indicated that the interface showed users that retrieval had occurred but not whether guidance came from aclinically reviewed source, suggesting that process transparency and source attribution are not interchangeable. As a formative study with a modest sample, a single site and no controlled comparison, this work establishes feasibility rather than effectiveness.