Streaming support for openAI's online requests

[lyuwen]

As addressed in #692 , the PR adds stream response support for openAI's online request processor.

Closes #692

[kartik4949]

@lyuwen Hello, thanks for the pr, could you let me know how to reproduce this?
which model and provider you used?

we have a similar setup and using longlived_response to flag deepseek model, where we are concatenating empty responses (deepseek sends these responses to maintain the connection) and get the reasoning response as joined.

[lyuwen]

I'm using a third party provider (alibabacloud) that hosts DeepSeek-R1 model. So the url check that activates the longlived_response would not work in this case. And it seems the MaaS provider does not send empty responses in non-stream mode and I would get a lot of timeout errors. After discussing with them, they recommended us to turn on stream.

[kartik4949]

@lyuwen got it makes sense, ill try to update it the code, however I need to somehow reproduce this as an integration test.
is it possible for you to check integration test and add one?
we use VCR, once you create the cassette file, you can redact all keys from the file.

let me know if this is not possible, Ill need to find other solution meanwhile for testing.

[kartik4949]

@lyuwen
https://github.com/lyuwen/curator/pull/1/files
please merge this into your branch and we should be good, ill create a follow up pr for tests.
Thanks

[lyuwen]

Thanks. I've merged your pr. And I'll take a look at testing.

[lyuwen]

I've just uploaded a cassette file for reproducing the case.

[lyuwen]

I just added a test that uses the mock reasoning dataset.

[kartik4949]

@lyuwen could you check makefile in this repository
before commit just do this.

make lint
make check

[lyuwen]

@lyuwen could you check makefile in this repository before commit just do this.

make lint
make check

After I ran these commands. There are a few files I haven't touched got changed. I only kept track of the ones that I modified.