Answers are grounded in this paper and cite their sources.
Archival research on human rights history records in Spanish, such as historical documents, faces the challenges of noisy Optical Character Recognition (OCR), fragmented testimony, and unstructured text that impede reliable retrieval. We implement GraphRAG, an automated framework that couples retrieval-augmented generation with knowledge graph construction to turn noisy corpora into interactive, verifiable network representations. To validate our proposed method, we apply GraphRAG to historical files on the assassination of Orlando Letelier, a well-known case. The pipeline cleans the OCR text, segments it, and extracts entities and relations under a bespoke ontology; it then performs entity resolution, full-text and vector indexing, community detection, and produces grounded summaries that cite explicit graph evidence. Across six retrieval strategies—vector search, a hybrid vector and keyword method, Cypher property queries, and entity-enhanced variants—we evaluate answers produced by GPT-4o mini and o3 mini against a historian-curated gold standard. The results show that Cypher queries that use properties dramatically reduce hallucinations, while hybrid methods improve recall, achieving 95% precision in a set of gold standard questions. Another case study on disappeared detained archives reveals dense personal and institutional hubs and illustrates GraphRAG's ability to reconcile conflicting testimony into coherent narratives. Our open-source framework offers a scalable blueprint for digital archives, museums, and human rights organizations, reshaping how societies preserve, interrogate, and govern historical data-intensive workflows.
No takes yet. Share an insight, caveat, or question.
Sanhueza et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: