Generative artificial intelligence, especially large language models (LLMs), could accelerate environmental analysis, but deployment is hindered by two gaps: limited structured domain knowledge and unclear strategies matched to environmental decision contexts. Here, this study constructs a textbook-based, China-centered environmental knowledge data set with hierarchical organization to enable reliable fine-tuning and benchmarking. Results show a consistent trade-off that fine-tuned models achieve modest gains in precision (+1%) and response efficiency (+52%) on standardized tasks but exhibit limited adaptability when embedded in agentic workflows (-3%). In contrast, state-of-the-art generalist models consistently outperform in system-level sustainability and interdisciplinary decision tasks (+10%), benefiting from stronger cross-domain reasoning and dynamic tool integration. Together, these findings support a layered LLMs' deployment strategy for environmental intelligence. Specifically, selective fine-tuning for stable, regulatory, and verification tasks, combined with agentic workflows anchored in up-to-date generalist backbone models for dynamic, data-intensive, and interdisciplinary decision-making. This work provides both a reusable data set foundation and a practical framework for deploying LLMs as scalable and reliable decision-support tools in environmental decision.