CEG: Claim-Centric Evidence Graphs for Scenario-Aware Agricultural Knowledge Recommendation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Agricultural knowledge recommendation differs from standard item ranking because the desired output is often an actionable field plan rather than a document, a passage, or an isolated tip. A useful recommendation must match the crop, growth stage, problem, environment, and safety constraints while remaining traceable to technical evidence. We propose CEG, a claim-centric evidence graph framework that represents agricultural guidance as scenario-level recommendation cards. Offline, CEG converts heterogeneous agricultural materials into source-grounded claims, claim and entity relations, local recommendation items, and scenario plans. Online, CEG retrieves scenario plans first, falls back to item-level and raw evidence only when needed, and reranks candidates with scene fit, evidence strength, plan completeness, safety, and state-prior signals. The final output is a structured card containing diagnosis, ordered actions, execution conditions, risk controls, evidence references, evidence gaps, and next actions. On a frozen four-domain auto-silver benchmark with 100 examples per domain, CEG Full reaches 0.9327 macro Hit@1 and 0.9725 average MRR@5. The strongest ablation, CEG without plan synthesis, reaches 0.8830 and 0.9429, while PageIndex is the strongest standalone comparator at 0.8227 and 0.8631, respectively. We emphasize that these ranking metrics measure the ordering of designated relevant candidates only; they do not establish that generated plans are complete, safe, or satisfy all agronomic constraints.