| Digital advocacy platforms like RiseApp play a crucial role in helping people find reliable information and avoid the overwhelming experience of endless social media scrolling. However, traditional search engines, relying on keywords or fuzzy matching, often fail to bridge the semantic gap between user queries and relevant advocacy materials. This paper presents an end-to-end Retrieval Augmented Generation (RAG) system that replaces lexical search with semantic neural retrieval. Our system identifies the top three pieces of advocacy content relevant to a user’s query. Along with those results, it generates persuasive text using affective messaging techniques, empowering users to share meaningful content and engage effectively with those who oppose advocacy efforts. To assess real-world readiness, we benchmarked six different search architectures by stress-testing them with 466 queries, some carefully written and some messy or informal, just like people actually use in practice. What we found is that neural Cross Encoders have a clear edge: they are much more robust, outperforming old keyword-based methods by 23% in Mean Reciprocal Rank (MRR), and they do a great job at making sense of everyday language, even when it is degraded or casual. Beyond retrieval, we evaluated the generation layer using an ensemble of RAGAS and DeepEval metrics to assess faithfulness and relevance. Our findings demonstrate that while the "Authenticity Narrative" approach increased hallucination risks, the "Reactance Avoidance" strategy achieved a high composite score (faithfulness and relevance) of 0.925 and 0.977. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.