Turns out you don't need a giant pile of retrieved text to teach a chatbot about Latin American culture - a knowledge graph gets you most of the way there.
Researchers benchmarked a graph-based approach called Graph-RAG against standard Retrieval-Augmented Generation on LatamQA, a multiple-choice dataset covering eight thematic categories tied to Latin American culture. They built the knowledge graphs straight from Wikipedia using an open-domain extractor called KGGen, with no manual cleanup involved. The graph-based system, G-Retriever, cut the base model's error rate by 72% using a standard knowledge graph, and by 78% when the graph was tuned toward the benchmark's own topics. That put it roughly on par with conventional RAG, and the gap closed further as the graph was steered toward task-relevant content.
This targets a real blind spot: language models are trained mostly on text skewed toward English-speaking, US and European sources, so facts about underrepresented regions get memorized poorly. A knowledge graph offers something plain-text retrieval doesn't - you can point to exactly which fact fed an answer, and update it without re-scraping the web. The trained component also transferred zero-shot to Portuguese, hinting this could scale across a language family without rebuilding the pipeline for each one.
But "built automatically from Wikipedia" cuts both ways - the graph inherits every gap and bias already baked into Wikipedia's own coverage of Latin America, which is the long-tail problem this project set out to fix.