Deepseek fails to call functions for streamed outputs

[igorbpf]

Question

Hi all,

My agent, which uses deepseek-chat model API (DeepSeek-V3.2-Exp), fails to call functions/tools when I try to stream the output. It works fine for sync/async output though. I was wondering if this is a model issue or a framework issue. I believe it's model issue since that works for OpenAI models. However, I was hoping we could do something about it since deepseek's API it's quite opaque and there is no reference to that.

Additional Context

No response

[DouweM]

@igorbpf That indeed looks like a model/API issue rather than anything we have control over in the framework: the request we send to the API is the same in both cases, except for the stream: bool argument.

When streaming, is the model trying to perform a tool call but it's not being recognized as such, or does it try to do something completely different from the non-streaming run, as if it wasn't aware of the tools at all?

[igorbpf]

Hey @DouweM

Thanks for the quick response. In fact the real issue was sitting between the chair and the computer. Nothing a quick look at the docs couldn't solve: the issue was that I was using run_stream instead of run_stream_events. Deepseek seems to be more verbose than others and this behavior fools the validation approach.

https://ai.pydantic.dev/output/#streamed-results