Fix memo bug, allow conversation rounds to run in parallel

[petersalas]

This change:

• Fixes a bug where memoized streams sometimes duplicated their content
• Refactors <Converse> to allow the rounds to run in parallel when the content allows

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Comments	Updated (UTC)
ai-jsx-docs	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Nov 20, 2023 7:03pm
ai-jsx-nextjs-demo	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Nov 20, 2023 7:03pm
ai-jsx-tutorial-nextjs	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Nov 20, 2023 7:03pm
ai-jsx-voice	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Nov 20, 2023 7:03pm

[farzadab]

I'm not sure I understand the difference 100% but I read it multiple times and I don't see an issue :)

[petersalas]

Yeah it's sort of mind-bending and I have to constantly remind myself why it matters. The relevant difference between:

yield AI.AppendOnlyStream;
yield x;
yield y;

and

return [x, y]

is that in the top one, x will (might, depending on what the parent does) be rendered to completion before y starts being rendered, whereas in the second one they will render in parallel. What makes it especially mind-bending is that in this case y actually has an internal partial dependency on x. So the trick is to construct a render tree with inter-branch dependencies rather than inter-frame dependencies where possible.