Abstract
Knowledge extraction from unstructured Earth Science documents into standardized knowledge bases is a complex task that used to be heavily reliant on manual curation and domain expertise. As an effort to automate this process, we proposed a modularized multi-agent system framework, Adaptive Geo Knowledge Extraction (AGeoKE), that leverages Large Language Models (LLMs) to automatically extract and standardize knowledge from geological documents while maintaining the flexibility to adapt to different document types and domains with minimum human intervention. The framework employs a Model Context Protocol (MCP) architecture with a series of LLM agent teams performing a streamlined sequential workflow from raw PDF Optical Character Recognition (OCR) preprocessing to final well-structured knowledge output with controlled vocabulary alignment. To address concerns about the quality of AI-generated content, the framework utilizes a series of quality control mechanisms, including multi-version reviewing, error reflection, and term matching validation, to ensure the reliability of the extracted knowledge and the robustness of the automated process. The proposed method features a schema-adaptive design that enables the framework to generalize across document types and domains through a configurable modular schema system. Case studies on the USGS Mineral Deposit Models and the NASA Lunar Sample Compendium, validated through manual expert evaluation, demonstrate the effectiveness and cross-domain generalizability of the proposed method, highlighting the significant improvements of the implemented quality control mechanisms. By combining automated workflow design, schema-driven knowledge extraction, and comprehensive quality validation, AGeoKE provides a scalable and adaptive foundation of knowledge extraction across Earth Science domains and beyond, enabling efficient transformation of legacy scientific documents into structured, machine-readable knowledge bases.