Simple is better: evaluating pipelines for medieval Latin manuscripts

In the field of natural language processing and computer vision, medieval Latin manuscripts represent an extreme challenge: calligraphic abbreviations, complex ligatures, degraded parchments, and archaic vocabulary with minimal resources. Recent research has evaluated different pipeline architectures for translating these texts directly from the image, and the main finding contradicts the current trend of stacking more and more components. The simplest systems—a specialized optical character recognition (OCR) model feeding a visual language model—clearly outperform configurations that incorporate retrieval-augmented generation (RAG) or post-OCR correction. This has been called the complexity paradox: each additional module introduces noise, context saturation, and error propagation that ultimately degrades the quality of the final translation. This principle is not exclusive to digital paleography. In the development of custom applications for businesses, it is often observed that adding processing layers, middleware, or artificial intelligence without clear justification increases computational cost, response time, and failure rate. The lesson holds true both for a manuscript pipeline and for a system of AI for businesses: before scaling up complexity, one must validate whether a direct and well-specialized approach solves the problem. In the study on medieval manuscripts, the Latin-specific OCR reduced the character error rate by up to 4.3 times compared to generalist models, despite having orders of magnitude fewer parameters. It is a reminder that specialization—like custom software—can deliver superior performance without the need for excessive infrastructure. At Q2BSTUDIO, we apply this philosophy when designing technological solutions: we start by understanding the client's specific domain, select the most appropriate tools, and avoid over-engineering. Whether developing AI agents for document automation, implementing AWS and Azure cloud services to scale processes, or integrating Power BI for business intelligence, we prioritize functional simplicity. Cybersecurity also benefits from this approach: a simple pipeline with fewer points of failure is easier to audit and protect. Therefore, when facing complex problems, the first question should not be 'how many technologies can we stack?' but 'what is the most direct and robust path?'. The complexity paradox invites us to reconsider our architectures, both in manuscript research and in the development of custom applications and artificial intelligence systems. Sometimes, simple is not only more elegant: it is more effective.