A condition-aware retrieval-augmented decision support framework for tomato cultivation management

Tomato cultivation management requires decisions that depend on growth stage, environment, and production scenarios. Retrieval-augmented generation (RAG) can ground large language model (LLM) outputs in external knowledge, but conventional retrieval often ignores such conditional constraints, leading to evidence that is topically relevant yet condition-inapplicable. This study develops TSCA-RAG, a condition-aware RAG framework for tomato cultivation question answering. TSCA-RAG extracts temporal, environmental, and contextual conditions from user queries using a structured label set and employs TSCAF-Retrieval, which combines semantic retrieval, BM25 keyword retrieval, and metadata-based condition retrieval through adaptive fusion and a cross-strategy consistency reward. A tomato cultivation knowledge base was constructed as fine-grained knowledge units with condition annotations and used to evaluate both retrieval and end-to-end generation. Retrieval performance was assessed using Recall@K, MRR, and NDCG@5, and answer quality was evaluated using similarity-based and rubric-based metrics under matched generation settings. On retrieval benchmarks, TSCA-RAG improves over Fine-tuned BGE-M3 with relative gains of 5.70% in Recall@1, 4.31% in Recall@5, and 4.76% in NDCG@5. In end-to-end evaluation, TSCA-RAG achieves higher Faithfulness, Correctness, and Utility, with an 11.29% increase in Utility compared with the strongest baseline RAG system. The condition extraction module attains an overall F1 of 81.8%, and a built-in confidence attenuation mechanism recovers approximately 53% of performance loss from single-dimension extraction errors. These results indicate that explicitly modeling cultivation conditions, combined with robust extraction and adaptive error mitigation, can improve evidence applicability and response usefulness for AI-assisted tomato cultivation decision support.