What are the Issues with NodeJS Code Formatting in PipeDream and GPT4, and How Can a Custom PipeDream NodeJS GPT aid it?

1

This topic was automatically generated from Slack. You can find the original thread here.

Hey everyone – been using PipeDream to build GPT workflows for about 3 months now. Biggest frustration with both Pi and GPT4 is that the NodeJS code is constantly formatted in a way that is not compatible with PipeDream (require vs. import, not calling step variables correctly, forgetting to add return objects and instead logging everything, etc.).

Built a PipeDream NodeJS GPT for myself that has saved me a tone of time, figured I’d share with the community: https://chat.openai.com/g/g-3ZRxztHre-nodejs-assistant-for-pipedream

It’s not perfect, but it’s the first time I’ve been able to get run-ready Node code with a single prompt.

2

appreciate the feedback. The code gen in the UI is formatted to do this, and I’m backporting that to Pi now! We learned a lot about getting reliable code output, formatted for Pipedream, with the code gen product. We have a lot going on here so backporting this to Pi simply slipped on the backlog but you’re right that you shouldn’t have to fight with it to get it working.

I’ll let you know when that’s done and maybe you can help test it for me.

3

I just updated the Pi code prompt. Before is on the left, after is on the right. Notice that it provides the correct app prop and accepts input for other vars via props.

Would love if people have time to stress-test this and send any issues you find!

Screenshot 2023-11-12 at 8.40.30 PM.png
Screenshot 2023-11-12 at 8.40.30 PM.png950×1784 192 KB

Screenshot 2023-11-12 at 8.40.07 PM.png
Screenshot 2023-11-12 at 8.40.07 PM.png984×1968 248 KB

4

thank you very much! Any plans on switching over Pi to use the Assistants API , GPT4-Turbo, and threads + code interpreter? I still find myself having to take Pi’s output into code interpreter to get more complex code, and the limited Pi context window means that after a few messages, it loses the original conversation and forgets what my objective is

5

Absolutely re: GPT-4 Turbo, as soon as the rate limits increase to the point where we can use it.

Re: the assistants API, not yet. Slack maintains essentially the same state and with the 128K context window we’ll have effectively unlimited history like with assistants. When we add Pi to the Pipedream UI and improve the backend, I am considering that though.

I’m curious exactly what you’re seeing re: Pi not generating complex code. Could you share some examples of what’s not working? If you’re seeing benefits I would absolutely consider plugging into code interpreter as a tool