Ideas for the next Playground

[Gyu07]

This document outlines preliminary ideas and considerations for the redesign of node architecture in Giselle. These concepts are in the ideation phase and have not yet been implemented or fully validated through technical review

Key Points

• [Development] Because models are evolving rapidly, immediate and flexible responses are required to accommodate expansions in API functionality.
• [UX] From a user experience standpoint, there is no significant difference in the usability barrier between classifying jobs as “Text-Gen” and classifying by API provider.
• [Marketing] Having a distinct color for each model’s node makes screenshots/frames of the Playground more visually appealing when sharing.

Ideas

• Create a Gen Node for each API provider (i.e., replace “Text-Gen Node” with “Model Node”)

  • Examples: OpenAI GPT-4 Node, OpenAI o1 Node, Claude Node, Gemini Node
  • Source handling can remain the same as before.
  • Each Gen Node may have different parameters and support for multimodal capabilities.

• Include WebSearch Node functionality in a Gen Node (it’s possible)

  • Example: Within the Gemini Node, toggle Grounding (Source: Google Search) On/Off.

Consideration

• Pros

  • It becomes easier to adapt Node functionality to the new features of each model (fully leveraging API performance without limiting features due to Giselle’s UX constraints).
    • Error handling and individual management also become simpler.
  • Users who have experience with multiple models will face a lower learning curve.
    • Before: Understand Giselle’s UX → Understand the Gen Node.
    • After: Understand the Gen Node → Understand Giselle’s UX.
  • The design can be varied for each Gen Node (model), making the interface more playful
    • Example: On ChatGPT’s 10th anniversary, highlight only the ChatGPT node, etc.

• Cons

  • The number of Gen Nodes increases (while individual management becomes easier, separate maintenance is required).
  • A re-architecture is necessary.
  • Although multimodal support becomes easier, prompts may need to be crafted carefully to extract specific outputs.
    • Example: If you only want images from the output, you might need an instruction like “Instruction: output only images,” or use like an ELT node, etc.

[shige]

ℹ️ This idea is currently being implemented in the following issue.

• Migrate from the current UseCase Node to the ModelProvider Node architecture #366