Enhancing Technical Knowledge Acquisition with RAG Systems: the TEI use case

Abstract: n an era dominated by an explosion of technical documentation across diverse domains, the need for effective knowledge acquisition mechanisms has become paramount. The assimilation of the “Text Encoding Initiative” (TEI) guidelines, for instance, presents challenges for organizations and individuals seeking to effectively adopt its encoding principles. Retrieval-Augmented Generation (RAG) systems emerge as a promising paradigm to address this challenge, seamlessly integrating information retrieval with natural language generation to facilitate the acquisition of technical knowledge from large documentation material. In this publication, we explore how RAG systems can mitigate these challenges while maximizing the benefits of TEI adoption, particularly in the context of learning and implementation. Challenges in learning and adopting TEI guidelines revolve around the complexity of the markup language and the diverse skill levels of users. Mastering TEI requires familiarity with its intricate syntax, encoding conventions, and domain-specific applications, posing a steep learning curve for novices. Furthermore, the extensive volume of its published guidelines poses another challenge, even for experienced users, in efficiently retrieving relevant information. RAG systems provide a novel approach to technical knowledge acquisition by seamlessly integrating the power of Large Language Models (LLMs) and specialized knowledge. The logic behind a RAG system lies in its ability to leverage pre-trained LLMs to generate informative and contextually relevant responses based on retrieved information. This approach has emerged to tackle the hallucination issue observed in generative models. It achieves this by enriching the context necessary for these models with knowledge sourced directly from relevant documents. Consequently, RAG systems offer a promising solution to TEI adoption and learning challenges by providing a more adaptive and interactive content. Through advanced natural language generation capabilities, RAG systems can generate tailored explanations, examples, and walkthroughs of TEI encoding practices, catering to the specific needs and skill levels of users. By leveraging retrieval mechanisms, RAG systems can retrieve relevant TEI guidelines and examples from its extended documentation, facilitating self-paced learning and knowledge acquisition. Such interactive systems can empower users by providing assistance in the creation of TEI-compliant markup, streamlining the encoding process, and thereby reducing errors and inconsistencies. Furthermore, RAG-generated summaries and explanations elucidate the rationale behind TEI encoding decisions, enhancing transparency and reproducibility in digital humanities research. Upon experimentation with state of the art models, we observed the persistence of some technical challenges for this ultimate goal. First, a pre-processing of the documentation is necessary to overcome issues related to the tokenization process. Moreover, a chunking strategy for such a rich documentation has to be carefully defined to enable more precise information retrieval and complete response from the AI assistant. In addition, choosing the right prompt remains crucial to frame the context and the expected outcome in order to generate accurate responses. In conclusion, RAG systems offer a transformative approach to learning and adopting TEI guidelines, mitigating challenges and maximizing benefits for knowledge acquisition. By leveraging RAG systems to facilitate TEI learning, organizations can empower users to unlock the full potential of TEI for data interoperability, scholarly communication and digital humanities research.