How e-p-armstrong/augmentoolkit works
The system architecture should be represented as follows:
Core Components (Top Level)
• [Pipeline Manager] -> Controls execution of pipelines
• [Config Manager] -> Handles configuration and settings
• [Engine Wrapper] -> Manages API/Model interactions
Pipeline Structure (Second Level)
• QA Generation Pipeline
• RPToolkit Pipeline
• Classifier Creator Pipeline
Each pipeline should show its internal steps as a flow:
For QA Pipeline:
• Text Input -> Chunking -> Question Generation -> Validation -> Answer Generation -> Output Dataset
For RPToolkit:
• Story Input -> Emotion/Feature Extraction -> Scene Generation -> Story Generation -> Rating -> Output Dataset
For Classifier:
• Text Input -> Rule Generation -> Classification -> Model Training -> Evaluation -> Output Model
Core Abstractions (Third Level)
Relationship between key classes:
• PipelineStep
• GenerationStep
• EngineWrapper
Data Flow (Overlay)
• Input processing flow
• API/Model interactions
• Output generation paths
• Configuration application points
External Integrations
Connections to:
• Various AI APIs (Together.ai, Fireworks, etc.)
• Local Model Servers
• File System (Input/Output)
• Training Systems (Axolotl)
Key Design Elements to Include:
Component Labels
• Clear, concise names
• Version/configuration indicators where relevant
• Input/output format specifications
Subsystem Boundaries
• Related components
• Isolation between pipelines
• Shared resources
Modular nature of the system
• Each pipeline is independent
• Common abstractions are shared
• Easy to add new pipelines
Resumability feature
• Intermediate storage points
• State management
• Recovery paths
Configuration flexibility
• Multiple model support
• API provider switching
• Custom prompt integration
Validation points
• Quality checks
• Error handling
• Fallback paths