Agentic LLM Framework Extracts Botanical Traits from Documents

Information retrieval and data extraction from complex documents remain challenging, particularly in specialized fields like botany where descriptive texts are rich but unstructured. This research introduces a novel agentic framework designed to embed and annotate descriptive document layouts, demonstrated through a plant science use case. The pipeline integrates several advanced techniques to achieve robust and interpretable data extraction. The framework begins by converting PDF documents into machine-readable text using Optical Character Recognition (OCR), followed by segmentation and indexing based on genus and species. Rule-based parsers then extract structured botanical traits. Crucially, ensembles of Large Language Models (LLMs) are employed to expand trait vocabularies and resolve ambiguities, enhancing the system's ability to capture fine-grained contextual and structural information. Applied to three regional botanical datasets, the system successfully extracted 55,737 trait annotations from 4,961 species, with LLM-based enrichment boosting total annotations by 59%, proving its scalability and reliability for large-scale botanical data extraction.