Clarification on how system nodes refine LLM-generated workflows

[ashfaq92]

Hi,
Thank you for making this data available — it’s been very helpful.

I have a question regarding how the system nodes operate to refine the LLM-generated generic workflows, as described in the paper.

From my understanding, the Thing Data (TD) provided to the LLM do not include URIs or other low-level implementation details, allowing the LLM to generate workflows at a high level. Then, during the deployment phase, it seems that the system nodes are responsible for refining these workflows by adding URIs and other specific details, making the final deployable workflows.

Could you please clarify how this refinement is handled automatically by the system nodes? Any code or more data would be appreciated. In addition, would you like to share your thoughts on the need of these system nodes in the presence of node-red-node-wot?

Thank you!

[warburec]

Hi,

Thank you for your interest in this work, it's exciting to see people taking a look at it.

You are correct, the system-nodes allow for a light weight descriptions to be passed into Node-RED (not requiring URIs and other low-level implementation details). In this work we didn't reduce the generated json before sending it to Node-RED (and the system-nodes), however we have since implemented this for future work.

As only the name was necessary in the work's current form I believe I manually ensured the Thing names were correctly added to the respective URI values in the generated workflow. This would be no different to having the system-nodes fill in the URIs automatically, I just didn't have time to automate this (I don't think I had to make many edits but I cannot confirm for sure). This ensures the system-nodes are loaded correctly based on the properties they expect, and can therefore link to WoT and deploy.

The system-nodes were a part of my earlier experimentation. I originally used them for an inconclusive (due to sample size) user study into to effectiveness of the system-level programming approach for IoT and continued to use them as the project's scope changed. Node-red-node-wot is a great extension which I aimed to compare against the system-level approach as it requires defining the specific Things and affordances. There is no reason why it cannot be used but an architectural component would have to be implemented to convert the generated workflows into the lower-level implementation required by node-red-node-wot before injecting it into node-red. The downside of using the extension is that inexperienced users will have a harder time understanding the implementation (visual-workflow in node-RED) as there are more options and URIs displayed to them.

I hope this provides enough clarification for you. If you need anything else don't hesitate to reach out!

Ewan

[ashfaq92]

Hi,
Thanks for the response. I have reached out to you on linkedin. We can have a chat when you are available!
Ashfaq