RAG as a Content-Analysis Assistant: Auditing SDG Discourse in Online Videos

Auditing how the Sustainable Development Goals (SDGs) are invoked in public communication is important for accountability, e.g., to detect superficial or strategic “SDG-washing”, but the currently most accurate and interpretable form of auditing – manual content analysis – does not scale to the volume and length of contemporary online video. Long-form online video transcripts are particularly challenging: they are noisy, weakly structured, and difficult to navigate while still requiring traceable evidence for coding decisions. We study retrieval-augmented generation (RAG) as an assistant to support SDG discourse analysis, helping analysts locate relevant transcript spans and draft codebook-style codes and claim/theme summaries with quotes. We instantiate this workflow on a French SDG-focused online video corpus and evaluate a set of practical pipeline variants and their qualitative contributions. Using a compact evaluation protocol, we find that query rewriting yields the largest and most consistent end-to-end gains. The main remaining bottleneck is provenance: models often produce plausible answers while blurring the boundary between retrieved evidence and background knowledge, motivating explicit guardrails in the analytical pipeline.