Compressor Blade Multi-Agent Design System — Core Prompts

This repository contains the core prompts of a LangGraph-based multi-agent system for compressor blade design. The system orchestrates six specialized agents through a hub-and-spoke architecture, with all domain knowledge encoded via prompt engineering on a general-purpose LLM (DeepSeek-V3) — no fine-tuning required.

System Architecture

                        ┌─────────────────────┐
                        │  Task Planning Agent │
                        │   (Coordinator)      │
                        └──────────┬──────────┘
                                   │
          ┌────────────┬───────────┼───────────┬────────────┐
          ▼            ▼           ▼           ▼            ▼
   ┌────────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐
   │  Blade     │ │ Perf.   │ │ Optim.  │ │ Physics │ │Knowledge│
   │ Generation │ │Predict. │ │  Agent  │ │ Valid.  │ │Synthesis│
   │   Agent    │ │  Agent  │ │         │ │  Agent  │ │  Agent  │
   │  (CDDPM)  │ │  (TeM)  │ │(GA/PSO/ │ │(CFD/FEA)│ │  (Q&A)  │
   │            │ │  (ViT)  │ │  LLM)   │ │         │ │         │
   └────────────┘ └─────────┘ └─────────┘ └─────────┘ └─────────┘

Prompt Index

Agent	Prompt	File	Section
Task Planning Agent	Task Decomposition	A1	A.1.1
	Runtime Conditional Gate	A1	A.1.2
	Intent Classification	A1	A.1.3
Blade Generation Agent	Parameter Extraction	A2	A.2.1
	Sensitivity Param Match	A2	A.2.2
Performance Prediction Agent	(tool-based, no prompt)	A3	A.3
Optimization Agent	Parameter Extraction	A4	A.4.1
Physics Validation Agent	Simulation Router	A5	A.5.1
Knowledge Synthesis Agent	Chat System Prompt	A6	A.6.1
	User Intent Analysis	A6	A.6.2

Key Design Decisions

• Conditional task preservation: Tasks guarded by runtime conditions (e.g., "optimize if not meeting targets") are always included in the plan; the runtime conditional gate (A.1.2) decides execution or skip based on actual results.
• Dual-threshold gating: Hard fail (< 95% of target) triggers auto-injection; borderline (95–100%) solicits user confirmation.
• File-state awareness: The simulation router injects current filesystem status of each scheme, enabling minimum-cost execution path planning.
• Two-pass parameter matching: Alias lookup → LLM semantic matching fallback for the 21-dimensional blade design space.

Citation

If you use these prompts in your research, please cite our paper:

@article{wu2026compressor,
  title={A Multi-Agent System for Compressor Blade Design},
  author={Wu, Yueteng and others},
  year={2026}
}

License

MIT License