New RAG system cuts OCR errors by 60% and boosts archive search accuracy 31%

A team led by Stergios Konstantinidis has developed an end-to-end framework that injects LLMs into historical archive processing. The system tackles two persistent challenges: noisy OCR (common in digitized historical texts) and rigid keyword-based retrieval. Its first component uses an LLM to refine raw OCR output, drastically reducing character error rate (CER) by 44.52% and word error rate (WER) by 60.95% on a dataset of 500,000 Swiss newspaper segments spanning 1762 to 2001. The second component implements a semantic retrieval pipeline with cross-encoder reranking, enabling users to ask natural-language questions via retrieval-augmented generation (RAG).

Evaluated across 384 test queries, the reranking pipeline improved NDCG@10 by 31.9% (from 65.99% to 87.05%) over traditional keyword baselines, with statistically significant gains in both answer correctness and context relevance. These results demonstrate that combining LLM-based OCR cleanup with semantic search can transform massive, noisy digital libraries into interactive, queryable knowledge systems. The work points toward a future where researchers can easily mine centuries of cultural heritage data without manual transcription or rigid search filters.