A Preliminary Exploration of the Knowledge Base for Excavated Documents Research

This article focuses on the application and transformation of artificial intelligence technology in the construction of knowledge bases for excavated document research. It systematically traces the evolution of excavated document knowledge bases from early digitized databases, knowledge graphs, and multimodal databases to new paradigms such as Retrieval Augmentation Generation (RAG), GraphRAG, Agentic RAG, and LLM-Wiki. This paper argues that while traditional RAG can improve the efficiency of document retrieval and question answering, it still faces limitations in research tasks unique to excavated documents, such as fragmented retrieval, insufficient context restoration, weak knowledge connections, and difficulties in version maintenance, particularly in areas like intertextual relationships, diachronic evolution, character shape comparison, historical interpretation, and cross-material connections. In contrast, the LLM-Wiki methodology proposed by Andrej Karpathy in 2026 emphasizes replacing temporary splicing during queries with incremental compilation, persistent maintenance, and structured connections, providing a new approach for knowledge bases to shift from "instant question answering tools" to "sustainably growing research-oriented knowledge systems." This paper analyzes the advantages and disadvantages of different tools, including IMA, WeKnora, MaxKB, and nashsu / llm_wiki, in terms of application threshold, knowledge organization, academic adaptability, and human-computer collaboration. The study argues that future construction of excavated document knowledge bases should adopt a fusion approach of "RAG instant retrieval + LLM-Wiki long-term accumulation," and further focus on multimodal understanding, structured memory, knowledge entropy control, and prevention of recursive distortion, in order to construct a digital humanities research infrastructure that combines traceability, evolvability, and academic rigor. In contrast, the LLM-Wiki methodology, proposed by Andrej Karpathy in 2026, emphasizes incremental compilation, persistent maintenance, and structured linkage over ad-hoc query-time stitching. This approac h offers a new paradigm for transitioning knowledge bases from "real-time Q&A tools" to "sustainable, evolving research knowledge systems." Furthermore, this paper analyzes several existing frameworks—including IMA, WeKnora, MaxKB, and nashsu / llm_wiki —evaluating their relative strengths and weaknesses regarding accessibility, knowledge organization, academic adaptability, and human-AI collaboration. The study proposes that future development of excavated documents knowledge bases should adopt a hybrid pathway of "RAG for real-time retrieval + LLM-Wiki for long-term consolidation." It further emphasizes the need to prioritize advancements in multimodal understanding, structured memory, knowledge entropy control, and the mitigation of recursive distortion. Ultimately, these efforts aim to establish a digital humanities research infrastructure characterized by traceability, evolvability, and academic rigor.