Problems when using toolcalls in streaming

[yp05327]

1. The return of acc.JustFinishedToolCall() - FinishedChatCompletionToolCall does not contain toolCallID, which is is required for openai.ToolMessage, so users need to record it by themselves.
2. We should append the origin tool call Message to the message list used in the API call.
   For non-streaming, you can simply append chat.choices[0].message like:
   https://github.com/openai/openai-[go/blob/952d231c9d5fce675262a6a639a99a62249d4a7f/examples/chat-completion-tool-calling/main.go#L62](https://www.golinks.io/blob/952d231c9d5fce675262a6a639a99a62249d4a7f/examples/chat-completion-tool-calling/main.go#L62?trackSource=github)
   For streaming, it seems that there's not implements about it, maybe something like openai.ToolCallsMessage? Or, simply use FinishedChatCompletionToolCall as the Message?
   Also, I could not find any examples about it. I only found this:
   https://github.com/openai/openai-[go/blob/952d231c9d5fce675262a6a639a99a62249d4a7f/examples/chat-completion-accumulating/main.go#L65-L68](https://www.golinks.io/blob/952d231c9d5fce675262a6a639a99a62249d4a7f/examples/chat-completion-accumulating/main.go#L65-L68?trackSource=github)
   If you don't append the origin tool call Message with only one openai.ToolMessage, you will get this error:

Invalid parameter: messages with role 'tool' must be a response to a preceeding message with 'tool_calls'.

[dlo]

@yp05327 Did you figure this out?

[dlo]

It's undocumented, but here's the solution for anyone else dropping in.

You just need to add an additional ChatCompletionMessage to your message list in the follow-up API call and a ToolMessage message.

Go playground link: https://go.dev/play/p/35kQ-Zx6jV2

And inline:

func main() {
	var ctx context.Context
	var system, user, functionName, toolCallId, arguments, result string
	
	client := openai.NewClient()
	stream := client.Chat.Completions.NewStreaming(ctx, openai.ChatCompletionNewParams{
		Messages: openai.F([]openai.ChatCompletionMessageParamUnion{
			openai.SystemMessage(system),
			openai.UserMessage(user),
			openai.ChatCompletionMessage{
				Role: "assistant",
				ToolCalls: []openai.ChatCompletionMessageToolCall{
					{
						ID:   toolCallId,
						Type: "function",
						Function: openai.ChatCompletionMessageToolCallFunction{
							Name:      functionName,
							Arguments: arguments,
						},
					},
				},
			},
			openai.ToolMessage(toolCallId, result),
		}),
		Model: openai.F(openai.ChatModelGPT4oMini),
	})
	defer func(stream *ssestream.Stream[openai.ChatCompletionChunk]) {
		err := stream.Close()
		if err != nil {
			panic(fmt.Errorf("error closing stream: %v", err))
		}
	}(stream)
}

[jacobzim-stl]

The tool call id has been added to the accumulator in the upcoming V2 release. Thanks for sharing a solution!

[yp05327]

@jacobzim-stl
Thanks for the work.
I found there are three branchs including v2:

which one is the 'main' branch for v2?

[yp05327]

@dlo
Which version are you using. I'm using 0.1.0-beta6, and I got this error:

cannot use openai.ChatCompletionMessage{…} (value of struct type openai.ChatCompletionMessage) as openai.ChatCompletionMessageParamUnion value in argument to append

Also, there's no openai.F in it.

[yp05327]

Ok, I figured it out.
The correct version is v0.1.0-alpha.67 in v2/next branch.

But in this version, openai.F is not exsist and you can not add openai.ChatCompletionMessage into []openai.ChatCompletionMessageParamUnion directly.
You should use openai.ChatCompletionMessage.ToParam() to convert it into openai.ChatCompletionMessageParamUnion