How adobe-research/dynasaur works
Main Components
a) Core Components:
• LLM Interface Layer (Azure OpenAI integration)
• Agent System
• Dynamic Action Generator
• Action Library Manager
• Embedding System
b) Supporting Components:
• Environment Handler
• Tools Collection
• Prompt Management System
Top Layer: Client Interface
Middle Layer: Core Processing
Bottom Layer: External Services & Tools
Key Relationships to Highlight
• User Input → Environment Handler
• Environment Handler → Agent System
• Agent System ↔ LLM Interface
• Agent System ↔ Dynamic Action Generator
• Dynamic Action Generator ↔ Action Library
• Action Library ↔ Tools Collection
Special Architectural Elements
Unique aspects:
• Dynamic Action Generation Loop
• Action Library Extension Process
• Embedding-based Action Retrieval System
Component Details
Key files to represent:
• dynasaur.py (Main System Controller)
• agents.py (Agent Implementation)
• actions.py (Action Management)
• prompts.py (Prompt Templates)
• scripts/tools/* (Tool Implementation)
External Dependencies
Connections to:
• Azure OpenAI API
• Azure Embeddings API
• Optional APIs (SERPAPI, GPT4V)
Implementation Notes
Annotations for:
• Python code generation capability
• Action composition mechanism
• Error recovery system
• GAIA benchmark integration
Data Flow Indicators
• → Solid: Direct calls
• ⇄ Double: Bidirectional communication