Check if we can use the most recent version of openai/openai-go

[mayabar]

Check if some of structures defined in openai-server-api can be removed and use structures from the recent openai/openai-go

[MondayCha]

I agree with this feature, I am testing my multi-agent application with llmd-d-inference-sim, but found that regarding ToolChoice, the latest openai-go already supports multiple types, and as a result, it causes llmd-d-inference-simerrors. I think switching to a structure in openai/openai-go will help achieve this

E1027 13:04:41.219268  601626 server.go:117] "failed to unmarshal request body" err="json: cannot unmarshal object into Go struct field ChatCompletionRequest.tool_choice of type string"
E1027 13:04:41.219299  601626 simulator.go:472] "failed to read and parse request body" err="json: cannot unmarshal object into Go struct field ChatCompletionRequest.tool_choice of type string"

[irar2]

@MondayCha Do you mean you are missing an ability to specify the tool you want to call in ToolChoice?

[MondayCha]

@irar2 Yes.
The current mainstream model supports the semantics of forcing the call tool: How to force models to call a tool This is important in multi-agent collaboration.

OpenAI has also officially updated the definition of ToolChoice, which no longer only supports string strings, but also supports custom tools: https://github.com/openai/openai-go/blob/ae042a437e4ebef4dffe088bf01d087ac94feaf2/chatcompletion.go#L2680

In my scenario, managed to solve the problem of not being able to parse the custom tool's ToolChoice in JSON in MondayCha@885aaac .

I wonder if this new syntax should be supported by the tool call options defined in openai-server-api first, or should it be switched directly to the structure in openai/openai-go?

[irar2]

@MondayCha At the moment we decided not to switch to the structures of openai-go. We may do that in the (maybe near) future, incorporating /responses API, but this will require a big code restructuring. So for now we are keeping our requests and responses as defined in openai-server-api.

Your implementation of ToolChoice looks good to me. It just needs to be continued in CreateToolCalls() function to actually choose the specified tool. Would you like to create a PR for this? If you only have code for the request parsing, you can create a PR only for it, and we will continue with the rest.

[MondayCha]

@irar2 Okay, I'm more than happy to, I expect to submit a canonical PR later today (Since I'm doing my research based on your simulator, thank you so much for such a good tool)