From Parametric Guessing to Graph-Grounded Answers: Building Reliable ChatGPT-like tools for Plant Science

Abstract

Large language models (LLMs) are increasingly used by plant biologists to summarize literature, generate hypotheses, and interpret experimental results. However, LLMs are unreliable sources of exhaustive, source-attributed facts, a critical limitation for the list-style queries that pervade plant biology (e.g., "list all transcription factors regulating secondary cell wall (SCW) biosynthesis in Arabidopsis"). Here, we query ChatGPT, Claude, and Gemini with such queries and demonstrate that none return complete gene lists with reliable citations. We trace these failures to how LLMs store knowledge: as statistical patterns distributed across billions of internal parameters, with no mechanism to guarantee completeness, provenance, or reproducibility. We also review fine-tuning mitigation strategies, including multi-task instruction tuning, parameter-efficient methods, and context engineering, that alleviate but do not resolve these limitations. We then discuss retrieval-augmented generation (RAG), which feeds relevant documents to the LLM at query time, and argue that while it improves source attribution, it remains impractical when answers require synthesizing information scattered across hundreds of papers. As an alternative, we advocate graph retrieval-augmented generation (GraphRAG), in which the LLM serves as a reasoning and language interface over a structured, provenance-linked knowledge graph (KG) that returns complete result sets reproducibly. We outline a practical GraphRAG architecture and survey existing plant KG resources. Finally, we discuss open challenges, including entity disambiguation, relation normalization and evidence grading, and propose a roadmap for building open, continuously updated plant KGs that can turn "read 1,000 papers" into a single reproducible query.