AgriEvoKG: Empirical Failure Modes and Remedies in LLM-Based Cross-Media Agricultural Knowledge Graph Construction

Large language models (LLMs) make it possible to build knowledge graphs from professional literature at scale, but how such graphs fail at that scale is rarely measured. We built \sys{}, a cross-media knowledge graph of swine, aquaculture, tomato, maize and rice production, and analysed its 199,418 entities and 277,053 relations together with the web system that serves it. An LLM extracted typed triples from the professional literature of each sector; 74.9% of the relations carry an evidence record of document, page and excerpt, and a multimodal embedding model links 25,158 entities to images and 14,132 to video clips. We identified five failure modes of construction. First, the sector schemas diverged: the two sectors with free relation labels hold 23,727 and 19,139 distinct labels. Second, in three sectors most numeric facts lost the name of the measured quantity or object, because 85.4% to 98.0% of their value triples use a generic relation. Third, evidence anchoring was asymmetric: the LLM copied 87.5% of the values verbatim from their excerpts but only 38.7% of the head entities. Fourth, exact-name fusion merged 6,805 entity names across sectors but only 7 relations, and about half of the cross-sector relations touch value-like names. Fifth, media links concentrated on hub items: 104 of 15,314 media items held 71.6% of the links. On the deployed system, 62.5% of the relations added from 50 new passages used new relation types, and in a case study on 100 photographs graph context confirmed the model's initial identification rather than correcting it. We organise these findings into a taxonomy that pairs each failure mode with its root cause, a remedy and a test, and argue that most of them cannot be seen in the small samples on which construction methods are usually judged.