This paper describes outcomes and challenges in human-scale document processing. We discuss a workflow that begins with document preservation, moves through text recognition, and ends with a catalogue that demonstrates capabilities of LLMs for research and archival work, while remaining attuned to the vision of research partners.
Until 2022, the Istmina Circuit Court archive, with documents from the 1870s to 1930s, was rotting, disorganized, and in garbage bags. Yet, this archive is a crucial source of Afro-Colombian history in an often-marginalized region of the Chocó in Colombia. In 2023, seven young people from Istmina and Quibdó worked with the Muntú Bantú Foundation, a community center focused on Afro-diasporic memory. With researchers from various universities, they were able to digitize the archive, which is available online at the British Library. While the project was successful, the digitization has enabled new workflows to catalogue and interpret the archive. This paper explores these workflows.
Throughout, we are interested in a key problem of equity in knowledge production: How can new tools be used to the benefit of local knowledge-producers? Our paper focuses on the work of cataloguing archival materials, a first step in enabling local researchers (and others) to make meaning. Project interns catalogued 330 Case Files and wrote a book of micro-history. Yet, the task of cataloguing 470 more cases remains daunting. We reflect on machine learning pipelines to extract text, catalogue the archive, and understand 61,000 images using Weasel, a digital workflow system.
To extract text, we built a workflow which fetches images hosted on the British Library; uses Kraken to segment text in each image; deploys Google Vision to extract handwritten and typewritten text; sends the image, the polygonal representations, and text to eScriptorium, where users can review it and correct the text. Here, we discuss both positive outcomes and ongoing challenges in recognizing text in typewritten versus handwritten documents, in segmenting images into regions, and working with these tools.
From there, to create a catalogue, we built workflow that downloads transcriptions from eScriptorium; uses spaCy named entity recognition to extract and link the names of people, places, dates, events, organizations; and employs open-source large-language models running locally via Ollama (mistral:instruct and mistral:8x7b) to generate summaries, timelines, catalogue entries, and other catalogue material. We run experiments with RAGatouille, ColBERT, LangGraph powered agent-based workflows to do further work on the archive. Finally, we export this material as a catalogue of extracted text, summaries, named entities, keywords, etc. into a fichero, a box of linked and tagged digital index cards in Markdown format which Obsidian and similar tools make accessible to non-technical users. Additionally, we use Nomic’s Atlas to map the data and metadata. The collection can be published to the web using 11ty or other static generators.
Here, we discuss steps to create a machine-generated catalogue, challenges to choose the right approaches, the costs of online commercial AI models, and the importance of the right tool.
