Case study demonstrates that tool integration elevates language model scheduling accuracy to 90% in shipyard operations, indicating tool access outweighs context capacity.
Key Points
To develop an autonomous on-premises large language model agent framework that resolves context degradation and functional separation in large-scale industrial project scheduling.
Orchestrated an agent system via LangGraph applying the reasoning and acting (ReAct) paradigm to couple plan generation with analytical feedback.
Combined local and global search across a schedule knowledge graph using graph retrieval-augmented generation (Graph RAG) alongside dedicated tools for schedule-wide analysis, modification, and optimization.
Evaluated system accuracy against tool-free open-source baselines and a long-context commercial model (Gemini) using real-world engineering project data from a shipyard.
The agent framework increased the scheduling accuracy of open-source backbone models to approximately 90% on the large-scale schedule.
Tool-free open-source models failed to fit the schedule within their context window, while a tool-free commercial model with extended context (Gemini) attained only 41% accuracy.
Component-wise ablation identified the absence of execution tools, rather than context window limitations, as the primary operational bottleneck.