This project originated in an earlier study of regional economic history with a focus on the trade of Nordisk Fjerfabrik, a Dainsh feather manufacturing company engaged in transnational commerce. The analysis draws on the company’s trade records preserved in the Danish National Archives, mainly from the 1910s and 1920s. As these materials have not been systematically digitized, access was limited to continuous photographic capture of documents within a restricted archival timeframe, resulting in a compact but information-dense set of archival images. This mode of rapid acquisition under constrained access is not exceptional, but reflects a long-standing practice among historians working across disciplines (Milligan 2020: 602) (Rutner, Jennifer / Schonfeld, Roger C. 2012). Under current technological conditions, researchers increasingly rely on digital processing to replace parts of traditional manual extraction and comparison workflows. However, archival materials collected in this manner often exhibit a high degree of heterogeneity in image quality, layout, and language, making them difficult to process using standardized workflows.

Corporate archives of this kind are common in regional and transnational economic history, where analytical value depends on systematic comparison across documents and over time. While digital humanities research has developed mature digitization and analysis pipelines for archival materials, including OCR, transcription, automated analysis, and more recently LLM (Large Language Model)-assisted workflows, recent work has increasingly explored the use of multimodal LLMs as alternatives to or extensions of traditional OCR systems, suggesting that they can, under certain conditions, outperform conventional approaches in both recognition accuracy and layout interpretation, even without extensive preprocessing or domain-specific training (Levchenko 2025: 725) (Xie et al. 2025) (Greif et al. 2025). At the same time, these approaches introduce new challenges, including hallucination, instability when processing visually complex or extended documents, and substantial computational cost when applied at scale. As a result, rather than fully replacing existing pipelines, current research increasingly explores how LLM-based processing can be integrated into practical workflows that balance accuracy, reliability, and cost under real archival constraints. In response, this paper presents a workflow-oriented approach that integrates these capabilities within a cost-aware and constraint-driven processing pipeline for archival collections of similar scale and conditions.

In the case addressed by this project, the primary challenge lies not in ways to further improve recognition accuracy, but in building a sufficiently complete and reliable digital archival structure at the initial processing stage, so that the materials can be reused and extended in subsequent research stages without repeated reprocessing. Starting from this research scenario, the present study examines how existing tools can be organized, under constrained access and resource conditions, to support a one-time, research-driven archival digitization workflow. From a methodological perspective, the project responds to these conditions by designing a reproducible and extensible archival processing workflow.

The overall pipeline consists of three stages: image preprocessing, OCR recognition, and OCR post-processing. As part of this design, layout segmentation and the annotation of key metadata are deliberately incorporated into the image preprocessing stage so that they can function as structural cues for subsequent cascaded use of LLMs. This staged design reflects an explicit attempt to align technical processing steps with the epistemic requirements of historical analysis and can address the analytical bottleneck of repeated manual comparison and enable historical questions to be formulated and tested at scale under constrained archival conditions.

At the image preprocessing stage, variations in image quality are treated as a foundational condition of workflow design rather than as exceptions. Based on an initial assessment of sample images, basic correction and enhancement techniques, such as cropping, deskewing, grayscale adjustment, and binarization, are applied adaptively. Human oversight remains an integral part of the workflow: the researcher first conducts an overall review of the images to assess processing difficulty and to distinguish materials suitable for automated handling from those requiring manual intervention. With the assistance of scripts, a metadata file is then created to record key information needed for subsequent stages, with particular attention paid to marking document continuity so that multi-page materials can be processed as coherent units.

At the OCR recognition stage, instead of adopting an end-to-end automated solution, this project combines local OCR engines with LLMs. Local OCR tools such as Tesseract and PaddleOCR offer stability and reproducibility for Chinese and English texts, but yield relatively high error rates when confronted with early twentieth-century Danish, mixed-language passages, or complex layouts, making them insufficient on their own for research purposes. LLMs, by contrast, demonstrate advantages in correction and refinement, but their use requires careful consideration of both time costs and API-related expenses. In response, the project adopts a hybrid strategy of local OCR followed by LLM-based correction, in which the output of local OCR provides a constrained candidate text space that guides LLM correction. This approach allows costs to be controlled while reducing the risk of hallucinations associated with fully end-to-end LLM-based recognition.

At the OCR post-processing stage, the corrected texts are routed to low-cost translation models to produce working texts suitable for research use. Throughout the workflow, automated scripts monitor execution status, checking for missing files or failed intermediate steps. The metadata file is continuously updated, with selected automatically extracted information organized as keywords or annotations to support subsequent analysis. On this basis, cascaded LLM calls can be introduced as needed to further process the curated texts or extracted features.

The project demonstrates that, in research contexts of this kind, workflow design is often more consequential than the performance of any single tool. By foregrounding human judgment and explicitly recording document continuity and structural cues in metadata, OCR and LLM components can be embedded within a processing pipeline that is inspectable, interruptible, and reversible, rather than deployed as opaque end-to-end automation. As a result, archival materials can be organized into a searchable and extensible intermediate structure after initial processing, providing a foundation for subsequent text analysis or database construction.

In this sense, the paper does not seek to propose a universal method for archival digitization, but rather presents a practice-oriented pathway aligned with the research lifecycle and shaped by concrete archival constraints. Under conditions of limited access and constrained resources, controlled automation and clearly defined workflow boundaries can help transform one-off digitization efforts into durable research infrastructure. This experience offers a practical reference framework for digital humanities research that relies on physical archives and operates under similar constraints.