An Interactive Language-based Artificial Intelligence Agent for Predefined Environmental Workflows: A Smart Farm Case Study

Engineering notes

Engineering notes will be added by the aipentium editorial team.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为大语言模型、生成式AI、ChatGPT相关技术、计算机视觉、深度学习等高价值论文补充中文说明。

Original abstract

This record contains the software, datasets, and example outputs supporting the manuscript: "An Interactive Language-based Artificial Intelligence Agent for Predefined Environmental Workflows: A Smart Farm Case Study" The system uses two agents — a diagnostic agent and a resource-advisor agent — built on a design that separates computation from interpretation: a predefined backend executes a fixed analysis pipeline, while a large language model explains the numerical outputs only (it does not select analyses or generate its own numbers). Contents (see README.txt):- 1_SmartFarm_application_Windows: a runnable Windows application (SmartFarm.exe) with its full Python backend source, embedded runtime, trained models, and the four farm-season datasets (Geumsan/Banseong, seasons 1-2).- 2_LLM_prompts_and_inputs: English/Korean documents reproducing the verbatim inputs sent to the LLM.- 3_agent_results: example diagnostic and resource-advisor run outputs (CSV tables, charts, reports, LLM interpretations).- 4_LLM_only_evaluation_responses: raw responses from an LLM-only baseline (ChatGPT, Claude, Gemini, Qwen3:8b), three runs per model. Software environment: Python (bundled runtime), Windows; default LLM backend Qwen3:8b served locally via Ollama, with optional ChatGPT, Gemini and Claude cloud fallbacks. Abstract:Environmental and agricultural analyses often rely on heterogeneous, data-limited datasets and domain-specific procedures. LLMs can assist, but autonomous method selection, code writing, prediction, and explanation can introduce unverified procedures and unstable outputs. To our knowledge, we propose the first AI agent architecture separating computation from interpretation. A predefined backend performs procedures and computations, while an LLM explains outputs, identifies caveats, and answers questions. This separation ensures reproducible computation and prevents the LLM from selecting unverified analytical procedures or generating its own numerical results, although interpretation errors may remain. As proof of concept, we developed diagnostic and resource-advisor agents for smart farm paprika yield analysis. The diagnostic agent evaluated cross-farm yield-prediction models, while the resource-advisor agent explored per-harvest-day resource-reduction candidates aimed at maintaining yield. Outputs were presented for review, not as validated predictions or operational setpoints. Compared with the AI-agent condition (1.00 for both metrics), LLM-only averaged 0.35 and 0.34 for methodological information coverage and explanation completeness, while ML-only scored 0.00. Predictive performance was low (train/test R2: 0.81/0.19), indicating more data are needed. This approach may extend to domains with standardized procedures or decision rules.

Links and sources

Comments