Inconsistency in generative UI streaming on Vercel Server. Sometimes works and sometime doesn't

[aa-wong]

I'm running into this weird issue on a vercel server with a pro plan. This implementation works very well and is super consistent on my local machine but when deployed on vercel server with a function timeout of 15 minutes. The yield animation works some times. Most of the time it doesn't work at all, with no animation indicator and sometimes it only after the user logs in using NextAuthJS which has no effect since the UI and actions.tsx is the same logged in and not logged in.

My implementation is the following dependencies:
NextJS: 14.1.4 (using AppRouter)
ai SDK ^3.0.17.

Within my action I am using render in the following code:

import { render } from 'ai/rsc;

const ui = render({
        model: 'gpt-4-turbo',
        provider: llm,
        initial: <AILoader message='Thinking...' />,
        messages: [
            {
                role: 'system',
                content: systemPrompt
            },
            ...thread
        ],
        text: async ({ content, done }) => {
            const message = {
                role: "assistant",
                content
            };

            if (done) {
                thread.push(message);
                updateState(aiState, { messages: thread }, session);
                return Renderer(thread);
            }
            return <AILoader message='Thinking...' />;
        },
        tools: procurementToolsActions(thread, aiState, session, updateState, systemPrompt)
    });

The Renderer function is another { render } from 'ai/rsc daisy chained to have dynamic generative UI response in accordance to the initial LLM response. We're supposed to be launching soon and everything works fine on my local machine but I'm getting a lot of inconsistent behaviour when deployed on the vercel servers. The streaming UI yield is not working sometimes and sometimes it does. Is there a way to have this more stable/consistent behaviour on the vercel server or at least indicate what the issue is? It's a little weird to have it work some of the time.

Here is a video of the behaviour that should be happening
https://github.com/vercel/ai/assets/14296176/9dd642b1-aaec-4e01-a7ea-bab63eab972c

This is a video of the behaviour when hosted on a vercel server
https://github.com/vercel/ai/assets/14296176/9a4d263b-2950-4e22-b219-106d033a9f24

[lgrammel]

Please check out these 2 troubleshooting pages:

• https://sdk.vercel.ai/docs/troubleshooting/common-issues/streaming-not-working-on-vercel
• https://sdk.vercel.ai/docs/troubleshooting/common-issues/timeout-on-vercel

Btw, we have deprecated render and recommend using streamUI.

[aa-wong]

Yes, I have tried the 2 suggestions you have provided and it still doesn't seem to be working properly. Is there any others areas we should look into or might indicate what might be the potential issue?

I've implemented 'force-dynamic' on the pages, the actions and the components. Nothing seems to be a resolve.

Also we are aware of the deprecation, it's scoped in our road map but right now in the process of release so we'll have to do this afterwards.