Retrieval-Augmented Generative AI for Analyzing Entrepreneurship in Africa

Systematic knowledge about entrepreneurship in Africa and other emerging economies is constrained by a structural data problem: the processes through which founders build high-growth ventures are richly documented in informal digital sources, including YouTube interviews, podcasts, and recorded conference appearances, but these sources have been unanalyzable at scale until now. This paper presents a validated, scalable methodology based on retrieval-augmented generation (RAG) and large language models (LLMs) for reconstructing founder journeys from fragmented multimodal content. We apply the methodology to the 16 founders of Africa’s six unicorn ventures as of 2024, processing over 1,000 YouTube videos and 9 hours of podcast audio into more than 30,000 embedded text chunks. The pipeline integrates multi-query expansion, semantic retrieval, structured narrative generation, and a refinement stage that iteratively improves factual grounding and removes unsupported claims . It is evaluated using precision, recall, F1, and faithfulness. We introduce the Context Completeness Score (CCS), a novel metric that assesses the causal and thematic depth of generated narratives beyond standard faithfulness measures. Results demonstrate high-fidelity narratives with faithfulness scores of 0.94 to 0.99. A key finding is that narrative quality depends more on thematic diversity than on data volume: reliable narratives are achievable with 800 to 1,200 well-curated text chunks. The methodology enables systematic analysis of founder journeys in data-sparse contexts, opening a new empirical pathway for entrepreneurship research in Africa and other emerging economies.