Historical archives pose a tough retrieval problem for retrieval‑augmented generation systems: documents are often OCR‑degraded, span heterogeneous genres and sources, and scholarly or institutional use demands strong source traceability. To address this, we introduce TRACE, a training‑free agentic retrieval framework focused on accountable source discovery over historical corpora. TRACE was built within the DECIDON project, which studies the circulation of political discourse between parliamentary debates and the press during the French Third Republic, involving digitised collections and institutional use cases. The prototype is internally deployed and currently serves 24 researchers across six partner institutions. Evaluation on the HistoriQA‑ThirdRepublic benchmark—1,752 French historical questions about 1887 parliamentary debates and newspapers derived from the Bibliothèque nationale de France digitised collections—shows TRACE achieving R@10 = 0.856 and MRR = 0.653, outperforming sparse, dense, graph‑based, and other agentic RAG baselines, with the largest gains on multi‑hop and cross‑corpus queries. Under the default hosted inference configuration the cost is about $0.02 per question, keeping the solution economically feasible for heritage institutions, labs, or companies lacking costly local GPU infrastructure. These results suggest that for large digital libraries and archives, retrieval accountability and corpus‑aware agent design can provide a practical alternative to heavier training‑based or graph‑construction approaches.
Review