feat (rsc): multiple tool calls for streamUI

[gclark-eightfold]

Summary

Adds the ability to invoke multiple tool calls for streamUI.

Addresses #1894, #1532 (comment)

NOTE: For usage with aiState, it's important to call aiState.done(...) only in the onFinish callback (if expecting multiple tool calls). -- Not sure if this is a breaking change.

Tasks

• rebase onto feat (rsc): add streamUI onFinish callback #1920 once merged to pickup refactored onFinish callback changes
• implementation
• test
• example
• docs
  • reference
  • example
• changeset

[bneigher]

#1894

[gclark-eightfold]

Is it possible to have text intermixed with tool calling?
I was able to get text before tool calling, but trying to prompt at least the OpenAI agent to require a text response after tool calling did not seem to work.

If text nodes can only appear once and before tool calling, this can be simplified instead of trying to keep track of nodes being written.

[gclark-eightfold]

Additionally, I noticed that with tool calling can come with a delay before the first chunk comes in for reader.read(). This had the unexpected effect of leaving out a last word or words of a text body that comes before tool calling and only shows up once the tool calls start streaming in.

[gclark-eightfold]

@lgrammel LMK if you have any thoughts on this implementation.
You had mentioned previously that streamUI would be refactored eventually to match streamText / other patterns, so let me know if you think supporting this would be best to do after that change or before.

I've ad-hoc tested this in my own application and it does seem to work, though I've made a few comments on the implementation where things were maybe not as ideal.

[dutsik-p]

Hi Guys!
Any news on releasing this feature.

[syedgaian]

Hi guys any news on , when will this feature be released?

[gclark-eightfold]

Hi guys any news on , when will this feature be released?

Haven't put this PR up for review because it's a bit buggy after updating it to the latest AI SDK version.
Currently getting these errors in my application I've patched with the changes in this PR:

• Unexpected Fiber popped.
• Detected multiple renderers concurrently rendering the same context provider. This is currently unsupported.
• TypeError: Cannot read properties of null (reading 'children')

[gclark-eightfold]

@lgrammel I reverted #2068 locally and tested on my application code and it fixed the issues described above.

Any idea how best to resolve this issue?

[artokun]

Ah, I was wondering why my app kept hanging up. This would be awesome to get released

[edspencer]

I had the same requirement and ended up building a layer on top of streamText to do it. It sounds like this will be natively supported in the SDK at some point, but for the time being I think it's still something one needs to put together oneself.

Check out https://github.com/edspencer/stream-multi if you are interested - the little video there shows the ability to have the LLM respond with both streaming text and any number of tools calls.

I published it as a npm module called ai-stream-multi. It basically has the same API as streamText(), but adds the concept of Segments, which are either a Tool or Text, and of which there can be any number in a given response from the LLM. Use the onSegment callback to update the AIState.

[zzh8829]

link all the related discussion, a new implementation could fix all these problems together.
#2915
#1894
#1895
#1976
#2915

[zzh8829]

I created a similar version here but using the new streamText api as the plumbing.
#1895 (comment)