Graph-based RAG system turns historical records into conversational AI

A new paper from Paula Font Solà, Adrià Molina Rodríguez, and Josep Lladós introduces a conversational retrieval system designed for historical digital libraries of penitentiary repression records. The system uses retrieval-augmented generation (RAG) to allow natural language queries, but goes further by building an on-the-fly knowledge model stored in a graph-based index. This graph acts as a long-term memory for the language model, capturing facts from expert archivists as well as insights derived from document retrieval. Over continuous use, the system can answer increasingly complex queries that require linking information across multiple documents or incorporating expert knowledge not present in the original text.

Accepted at ICDAR2026, the approach addresses a key limitation of standard RAG: the inability to interpret document collections holistically. By storing both explicit and inferred knowledge in a graph, the system enables richer responses, such as discovering connections between distant records or integrating contextual historical expertise. For archivists, this means moving from simple fact extraction to dynamic, context-aware analysis of large historical corpora. The system is particularly suited for sensitive collections like penitentiary repression archives, where expert interpretation is often crucial to understanding fragmentary or coded information.