1. Introduction

The Annals of the Joseon Dynasty are among the most extensive and systematically digitized corpora of premodern East Asian historical sources. At present, the Classical Chinese original texts of the Annals are available in XML format, structured by reign, date, volume, and entry, while modern Korean translations can also be accessed and searched through online services. This digital infrastructure is highly useful for retrieving specific reigns, dates, persons, events, and terms. However, it remains primarily a search-oriented environment for locating predefined textual units, rather than a “question-answerable research archive” capable of responding contextually to complex scholarly questions.

Recent advances in large language models have significantly expanded the possibilities of large-scale text processing and question answering. Yet general-purpose language models continue to face limitations when dealing with historical corpora such as the Annals, particularly in handling chronological context, official and institutional terminology, semantic differences between Classical Chinese originals and modern Korean translations, and the heterogeneous narrative structure of annalistic entries. Retrieval-Augmented Generation, or RAG, offers a possible way to address these limitations by grounding generated responses in external textual evidence. However, the performance of RAG for historical sources depends less on the choice of language model alone than on how the source materials are segmented, stored with metadata, retrieved, and reconstructed.

This study redefines RAG not merely as a response-generation technique or analytical tool, but as a research interface that mediates the way historians formulate and refine questions. Using the Annals of the Joseon Dynasty as a case study, it examines the technical and methodological issues that arise when designing RAG systems for large-scale annalistic archives. In particular, the study focuses on the fact that the Classical Chinese XML texts and the Korean translations are not provided in the same structured format, and therefore asks how the XML-based structure of the original text can be connected with the semantic accessibility of the translation. It further proposes chunking strategies, vector database design, original–translation linkage, and configurable retrieval pipelines as necessary components for transforming a search-oriented digital archive into a question-oriented research environment.

2. Research Background and Problem

Digital Humanities has developed a wide range of methods for accessing large textual corpora, including keyword search, index-based browsing, textual encoding, XML- and TEI-based structuring, and distant reading. These approaches have made major contributions to the preservation, access, exploration, and pattern-based analysis of historical materials. In the case of large historical corpora such as the Annals of the Joseon Dynasty, the structure by date, volume, and entry allows textual units to be identified with relative stability. In this respect, existing digital archives have greatly improved the accessibility of historical sources.

Nevertheless, existing systems are largely optimized for cases in which researchers already know the keyword, person, date, or event they want to search for. Such lookup-oriented interaction is effective for confirming specific terms, individuals, dates, or passages, but it is less suited to the process through which historians pose complex questions, explore related sources, and revise or expand their questions iteratively. Historical research is not simply a matter of finding one relevant entry. It often requires repeated comparison across multiple entries, attention to chronological change, variation in institutional terminology, and differences in the narration of events.

This study locates the potential of RAG precisely at this point. RAG can function not only as a search mechanism for large corpora, but also as an interface that connects scholarly questions with textual evidence. In the case of the Annals, which possess both an annalistic structure and an asymmetric relationship between Classical Chinese originals and Korean translations, the nature of the response changes significantly depending on which textual unit is retrieved and which linguistic layer is referenced. The central question of this study is therefore not simply whether RAG can generate good answers, but what kind of retrieval design is required to make a historical corpus such as the Annals of the Joseon Dynasty into a question-answerable archive.

3. Data Characteristics and Technical Challenges

The Annals of the Joseon Dynasty are an annalistic corpus compiled by date and entry. The Classical Chinese original texts provide relatively clear textual units and metadata through XML-based structuring. The Korean translations, however, are not provided in the same XML structure as the original texts. For this reason, building a RAG system requires addressing how the structure of the original XML data can be connected to the semantic information contained in the translations. This is not merely a matter of data collection, but a technical and methodological problem that directly affects retrieval units, metadata design, citation grounding, and response verifiability.

Chunking strategy is one of the most important issues in designing a RAG system for the Annals. Entry-level chunking has the advantage of preserving the original documentary structure of the corpus, but the length and density of entries vary considerably. Some entries are short factual records, while others include lengthy discussions, memorials, diplomatic documents, or explanations of institutions. If an excessively long entry is treated as a single chunk, retrieval sensitivity may decline. Conversely, if entries are divided uniformly by token length, the internal narrative structure and documentary context of the original entry may be damaged.

For this reason, a hybrid chunking strategy is required. The entry should serve as the default retrieval unit, while entries exceeding a defined length threshold should be subdivided into smaller units. In this process, metadata such as the entry identifier, date, reign, volume, and entry order from the original XML must be preserved. At the same time, if the Korean translations are used as auxiliary material for retrieval, they must be aligned with the original entries or at least linked through a common entry identification system. In this study, the original text and the translation are therefore understood not simply as two textual versions, but as asymmetric data layers that perform different functions in retrieval and verification.

4. System Design: RAG as a Research Interface

The system proposed in this study is a RAG-based research interface that supports question-oriented exploration of the Annals of the Joseon Dynasty. Its purpose is not merely to generate a single answer. Its more important aim is to build an experimental retrieval environment in which historians can adjust retrieval units, linguistic layers, metadata, and search scope according to the nature of their research questions. In this sense, RAG is understood not as a fixed pipeline, but as a configurable retrieval design space shaped by scholarly purposes.

First, chunking is implemented through a hybrid strategy. The entry of the Annals is used as the default retrieval unit, while entries that exceed a certain length are subdivided according to predefined criteria. This design seeks to balance documentary coherence with retrieval efficiency. It addresses the tension between preserving historical context and improving the sensitivity of vector-based retrieval.

Second, the vector database is designed not as a single fixed repository, but as an experimental space in which multiple retrieval conditions can be compared. Even when the same Annals corpus is used, retrieval results may vary depending on chunking strategy, metadata configuration, original–translation linkage, and embedding target. This study treats such variation not as an incidental problem or error, but as a key design variable that shapes the response quality of RAG systems for historical sources. The vector database is therefore constructed as research infrastructure for repeatedly testing and adjusting retrieval behavior.

Third, the retrieval pipeline is designed to separate but connect the roles of the original text and the translation. The Classical Chinese original is valuable for preserving original expressions, institutional terminology, and documentary citation. The Korean translation, by contrast, is useful for connecting modern natural-language questions with the semantic content of the source. Accordingly, the proposed RAG design uses the original XML as the basis for documentary structure and citation grounding, while using the Korean translation as an auxiliary layer for semantic retrieval and query understanding.

Fourth, the retrieval pipeline separates retrieval, filtering, reconstruction, and generation into distinct stages. The purpose of this structure is not to depend on a particular framework, but to clarify the role of each stage and preserve configurability. For example, combining vector search with metadata filters such as reign, date range, person, office, or institutional term can produce retrieval results better suited to historical research than semantic similarity search alone. In this respect, the core of RAG in this study lies not in the generation stage itself, but in the design of retrieval and textual grounding appropriate to historical sources.

5. Contributions and Future Directions

This study is an initial technical and methodological attempt to apply RAG to Korean premodern historical sources such as the Annals of the Joseon Dynasty. Its main contribution is not to claim the performance of a completed question-answering system, but to define explicitly the upstream design elements that determine whether historical sources can become answerable through RAG. Chunking strategy, vector database design, linkage between the original XML and Korean translation, and the configurability of the retrieval pipeline all directly affect the final response quality of the system. This study presents these elements not as fixed preprocessing steps, but as design variables that must be adjusted according to the goals of historical research.

The study also positions RAG as reusable backend infrastructure for historical scholarship. The annalistic structure, uneven entry length, and structural asymmetry between original XML texts and translations found in the Annals are not unique to this corpus. Similar problems may arise in other large-scale historical corpora. The design principles proposed here can therefore be extended to other premodern sources, including the Daily Records of the Royal Secretariat, the Records of the Border Defense Council, collected writings, and other state-compiled materials. Future research may expand this work through domain-specific embedding models, fine-tuned LLMs for Joseon historical language, reranking of retrieval results, and benchmarks for evaluating response quality.

In conclusion, this study proposes design principles for a RAG-based research interface that can transform the Annals of the Joseon Dynasty from a searchable archive into a question-answerable archive. It understands RAG not simply as a technology for generating answers, but as digital historical research infrastructure that changes how scholars formulate questions, revise them, and trace documentary evidence. Ultimately, this study argues that AI-based access to large-scale historical sources is not merely a matter of automation or convenience, but a methodological issue that reshapes the relationship between historical questions and archival evidence.