The GUI is developed using Flask/HTML for the front-end and Python for the back-end, ensuring cross-platform compatibility, usability, and responsive design. All technical details can be found in the GitHub repository and the dedicated Wiki documentation. To ensure user-friendliness, a graphical installer is also available for Windows and macOS. The proposed framework operates entirely locally, without reliance on external APIs or proprietary models, ensuring data privacy, reproducibility, and long-term sustainability. All experiments presented in this work were conducted on consumer-grade hardware: an Intel i7-12700K CPU, 48GB of RAM, and an NVIDIA RTX 3070ti GPU with 8GB of VRAM. The entire pipeline is also tested on a laptop and a MacBook Air M3 chip configuration.
Handwriting recognition evaluation demonstrates competitive accuracy with CER averaging 6.2% and WER at 18.6%. The CER of approximately 6% represents approximately 1 error per 17 characters, requiring correction but substantially lower than the cost of complete manual transcription. Such an error density, while requiring verification, transforms the task from exhaustive data creation to rapid validation.
Structured parsing performance reveals non-linear learning dynamics characteristic of in-context learning in LLMs. Across the few-shot continuum, three distinct performance regimes emerge. With 1 example, few-shot learning achieves 77.8% accuracy, 38.5% precision, 37.7% recall, and 37.9% F1, with EMA (Exact Match Accuracy) at 13.6% and PMS (Partial Match Score) at 66.6%. With 3 examples, performance jumps to 89.0% accuracy, 51.0% precision, 51.0% recall, 50.9% F1, EMA at 52.6% and PMS at 83.5%. With 5 examples, accuracy reaches 91.5%, EMA 60.4%, PMS 87.6%. With 10 examples, the system achieves 93.1% accuracy, EMA 67.6%, PMS 90.3%. The most dramatic improvement occurs between 1 and 3 examples, where EMA increases from 13.6% to 52.6%. From a workflow perspective, this reliability transforms the task from data creation to rapid validation, with high PMS indicating that most errors are isolated and easily correctable.
Processing time analysis of the complete pipeline on 50 sheets containing 240 drawings reveals approximately 103 minutes (1h 43min) total, yielding a throughput of 140 drawings per hour (2.33 drawings per minute). This aggregate figure masks important distributional characteristics. The time breakdown shows that automated components consume the majority of processing time, while manual filing and annotation represent the most time-intensive human-operated steps. Image extraction requires approximately 1 minute 13 seconds, OCR approximately 5 minutes 42 seconds, cleaning approximately 2 minutes 30 seconds, vectorization approximately 18 minutes, and layout generation approximately 27 minutes per 50-sheet batch. Manual filing and annotation each require approximately 13 minutes. These figures demonstrate that the automated pipeline reduces per-drawing processing time by approximately 65% compared to traditional manual workflows.
Usability study and perceived time savings
A formal usability study with five domain experts confirmed the framework's practical applicability. Participants reported a median perceived speedup of 40× over traditional workflows, with a range of 17.5×–120×. The System Usability Scale yielded an average score of 82, indicating strong usability. Qualitative interview themes highlighted most positive feedback regarding the speed and accuracy of automated scanning and layout generation, while the most common pain point concerned vectorization quality for complex decorative motifs. Feature requests centered on improved handling of cursive handwriting variants and expanded archaeological component taxonomies. The net promoter score was +42, indicating overall positive adoption intent.
Participants also estimated that the phase providing the greatest time savings was automated layout generation, followed by OCR-assisted text extraction. The phase needing most improvement was identified as vectorization of complex profile drawings. Task completion status showed that 4 of 5 experts were able to complete all assigned tasks without requesting assistance, with help requests concentrated in the vectorization and semantic parsing stages. The study confirmed that AI-assisted tools can substantially reduce cognitive load while maintaining or improving output quality in archaeological documentation contexts.
Strategic redistribution of cognitive labor
The paper addresses the strategic redistribution of cognitive labor toward augmentation rather than automation. The traditional post-production workflow places substantial burden on archaeologists for repetitive, mechanical tasks such as scanning, cropping, transcription, inking, and layout. PyPottery does not seek to eliminate the human role but to reposition it: the archaeologist transitions from executor of routine tasks to validator and curator of AI-generated outputs. This redistribution preserves the interpretive dimension of archaeological work while offloading mechanical burden to specialized models. The framework's open-source architecture ensures that the scientific community can inspect, modify, and extend the tools, facilitating collective and shared knowledge growth. By lowering adoption barriers through consumer-grade hardware compatibility and transparent model choices, PyPottery enables laboratories and research projects with diverse resource levels to integrate AI-assisted documentation into their workflows. The demonstrated time savings, combined with strong usability metrics, suggest that such tools can help mitigate the post-production bottleneck that currently limits the publication of valuable ceramic assemblages.