Agricultural Research Plant Science
Risk-Aware, Evidence-Traceable Advisory for Salt-Affected Soil Management Using Hybrid Retrieval and Agronomic Safety Constraints
Abstract
Safe agronomic advisory for salt-affected soils requires decoupling soluble salinity from exchangeable sodicity, alongside rigorous evaluations of irrigation-water quality, shallow groundwater dynamics, crop salt tolerance, and amendment feasibility. We developed an agricultural decision-support framework (AgriRAG) that couples hybrid document retrieval with a deterministic, multi-variable agronomic safety controller. The underlying knowledge base integrates 952 curated reference documents and a targeted safety layer of 66 verified nodes derived from 13 authoritative agricultural standards and handbooks. In standardized retrieval evaluations across six safety-critical domains, the architecture achieved 100% top-five reachability across 72 primary queries and 18 robustness paraphrases. To enforce physical and chemical constraints, the controller evaluates electrical conductivity ( $$\:E{C}_{e}$$ ) independently from exchangeable sodium indicators ( $$\:\text{E}\text{S}\text{P}$$ and $$\:{\text{S}\text{A}\text{R}}_{e}$$ ), halts provisional amendment calculations upon analytical conflict, and executes hierarchical safety vetoes against shallow saline water tables. Evaluated across a 360-case multi-variable benchmark, the system reproduced all 633 predefined functional assertions with 100% compliance, while blinded expert review confirmed significant safety gains over unconstrained baselines. This framework establishes an auditable decision interface enforcing deterministic constraints and verifiable evidence traceability for operational agricultural decision support.