From a backend perspective, this matches what I've seen building RAG pipelines. The issue is usually that standard chunking algorithms often sever semantic context in long-form text, even with overlap. Structured data or explicit Q&A formats survive retrieval much better because the embedding actually captures a complete thought rather than a fragment.