Architecture Node Types, Defaults, and Validation Warnings
How typed node definitions and property panel warnings keep visual model design safer in the Agentic Assistants architecture builder.
Visual Tools Need Guardrails
Drag-and-drop UIs can hide invalid configurations. The architecture tooling in Agentic Assistants avoids that by centralizing node schema in webui/src/components/model-architectures/types.ts.
Each node definition includes:
typelabelcategorydescriptiondefaults
That gives both the canvas and properties panel a single source of truth.
Node Schema As Product Contract
The current node types span IO, transformer blocks, and core layers:
input,outputembedding,positional_encoding,mha,ffn,residual_normlinear,conv1d,dropout,activation
Because defaults are explicit, newly dropped nodes are immediately usable and predictable.
Property Editing Design
ArchitecturePropertiesPanel.tsx handles configuration updates by key while preserving node-level metadata. It also adapts field type by default value:
- numeric defaults render number inputs
- structured defaults render JSON text areas
- selected enums (activation, positional mode) render dropdowns
This keeps editing fast without losing flexibility.
Warning Surfaces
The panel supports a warning list and each node supports a warning field. That matters because validation should appear where users are already editing, not in a separate diagnostics page.
Even simple warning surfaces improve architecture quality when teams iterate quickly.
Practical Takeaway
If you build a visual model editor, make schema and defaults explicit first, then add interactions. Strong typing in the node catalog is what makes export and validation reliable later.