Beyond Description: Expanding Access Through Visual Search
LUPA, An AI-Driven Multimodal Similarity Search Tool Developed by the Latvian State Archives of Audiovisual Documents
By Natālija Lāce
Archive Expert, Electronic documents Department
Latvian State Archive of Audiovisual Documents
National Archives of Latvia
Chair of ICA/PAAG
23 September 2026
The rapid growth of digitised and born-digital photographic collections in Latvian State Archive of audiovisual documents has transformed the possibilities of archival research. Yet this greater availability has also made one long-standing limitation increasingly visible: a photograph can only be found through conventional search if someone has first described what it contains.
Archival metadata is necessarily selective. A photograph may depict people, buildings, objects, technologies, landscapes or countless other details that were never identified when the document was described. These visual elements may be clearly present in the image yet remain effectively inaccessible to a researcher because they do not exist as description in the archives catalogue.
It was precisely this gap between what is visible and what has been described that led the Latvian State Archive of Audiovisual Documents to develop LUPA, a specialised multimodal similarity search tool for photographic collections. Drawing on artificial intelligence and OpenAI technologies, LUPA allows users to explore photographs through visual resemblance rather than text alone. It can identify similar images and faces, while archival metadata and chronological filters help place those visual connections within their broader documentary context.
Technical framework and data preparation
Developing LUPA required more than applying artificial intelligence to an existing digital collection. Before visual search could be introduced, the photographic documents themselves had to be prepared in a form that was technically consistent and reliably connected to their archival descriptions. The working dataset comprised 680,614 TIF and JPG files, representing approximately 7.54 TB of data.
The preparation process involved verifying file names against archival identifiers, removing duplicates and unsuitable files, standardising folder structures, and resolving technical issues such as incorrect image orientation and damaged files.
The reliability of visual search depends directly on the quality and consistency of the underlying archival data. Every digital image must remain correctly linked to its archival identifier and descriptive metadata throughout the entire workflow. Without this connection, visual similarity alone has limited archival value: a system may identify a relevant image, but the result must still lead the researcher back to the correct document and its documentary context.
Metadata preparation was therefore an equally important part of the work. 39,532 descriptive records were extracted from the archival database “Redzi, dzirdi Latviju!”. The selected data included archival identifiers, titles, photographers, annotations, places of photography, dates, years and persons represented in the photographs. These records were then transformed into a more consistent structure for use within LUPA.
Full-image and facial similarity search
LUPA supports two forms of documents retrieval: full image search and facial similarity search. Both are based on vectors, numerical representations of visual information that allow photographs and faces to be compared computationally. Separate vector collections are maintained for complete images and for faces; the current system contains 489,623 image vectors and 1,066,835 facial vectors.
The two search options serve different research needs. Facial similarity search can help locate photographs in which the same or visually similar persons appear, while image similarity can reveal recurring buildings, interiors, objects, technologies, landscapes and other visual elements across the collection. This functionality is exposed through the LUPA web interface, where users can search for similar images or faces, refine the results by year and archival metadata, and examine the retrieved photographic documents in a visual gallery.
Extending access to photographic documents
LUPA expands the possibilities of archival research by making visual content itself searchable. This is particularly valuable where photographic documents have only limited descriptive metadata or where relevant persons, objects, buildings or other visual elements were never identified in the archival description. Visual similarity search can reveal relationships across large collections, support the identification of previously unknown persons and bring forward connections that would be difficult to detect through human observation or keyword search alone.
At the same time, LUPA is designed as a research and decision-support tool rather than a system for automatic identification. A visually similar result indicates a possible connection, not a conclusion. Existing descriptions, dates, provenance and the wider historical context remain essential for assessing the relevance of the result, while archival identifiers provide a direct link back to the document and its context.
An important strength of LUPA is that it has been developed and implemented locally within the infrastructure of the National Archives of Latvia. Built around the archive’s own collections, identifiers, metadata structures and research practices, the system can be adapted as institutional requirements evolve. Local deployment also provides greater control over archival data, system architecture and future development, while reducing dependence on external providers.
For the Latvian State Archive of Audiovisual Documents, LUPA therefore represents more than the introduction of a new AI-based technology. It is a locally developed archival research tool that extends access to photographic documents while preserving the connection between visual discovery, archival context and professional interpretation.
