LLM-Powered Automated Information Extraction from Archaeological Reports: From Proof of Concept to Production-Grade Engine

The Korean heripo-lab team developed an LLM-based PoC project for automated metadata extraction from archaeological reports, and on this basis, open-sourced the production-grade engine heripo engine. This project addresses the pain point of difficulty in retrieving and analyzing unstructured information from PDF archaeological reports. It achieves structured extraction through an end-to-end pipeline, has published academic papers, and spawned a cross-domain technical ecosystem.

In the field of archaeology, a large number of excavation reports in PDF format contain rich unstructured information. Manual entry is costly and difficult to handle massive literature. The Korean heripo-lab team developed an LLM-based automated metadata extraction pipeline to address this pain point. Relevant results were published in the journal Heritage: History and Science, documenting the entire process from proof of concept to deployment.

The core process of the system is divided into three stages:

1. Document parsing and preprocessing: Extract text from PDFs (assuming they have a selectable text layer; scanned versions require additional OCR).
2. LLM-driven information extraction: Identify key metadata such as site names, ages, and artifact lists through prompt engineering.
3. Data standardization output: Convert to a unified format, laying the foundation for knowledge graph construction.

The team selected three representative Korean archaeological reports (Buyeo Hwajeosan Baekje Orchard Site, Jeju Hangpaduri Hangmong Site Inner Fortress, Gongju Seokjangri Paleolithic Site) for testing. Results show that the LLM solution can accurately identify key information and significantly reduce manual workload.

The heripo engine, open-sourced in January 2026, has been upgraded based on the PoC:

• Integrated Docling SDK to enable OCR for scanned PDFs;
• Native GPU optimization for Apple Silicon to support local processing;
• TypeScript monorepo architecture + 100% test coverage to improve engineering quality;
• Scalable data pipeline to adapt to the needs of different institutions;
• Compatible with multiple LLM providers (OpenAI, Anthropic, etc.).

The project's technical accumulation has spawned the LLM Newsletter Kit, which builds a type-safe AI news briefing engine. It supports the Research Radar service covering 62 data sources, automatically generating and distributing weekly briefings. The cost per briefing is $0.2-$1, with a click-through rate of 15%, demonstrating the cross-domain universal value of the core technology.

Insights:

• Deep integration of domain knowledge and AI is key;
• Phased development (PoC validation → production refactoring) reduces risks;
• Open-source collaboration accelerates evolution. Outlook:
• Integrate image understanding capabilities with multimodal models;
• Cross-language models facilitate cross-border archaeological data integration, providing infrastructure for global cultural heritage research.