fix(openai): default Azure LLM model to deployment

[nightcityblade]

Fixes #6209

Summary

• stop LLM.with_azure() from exposing the unrelated gpt-4o default when Azure callers only set azure_deployment
• use the Azure deployment name as the fallback model identifier so telemetry and llm.model reflect the configured deployment
• add a regression test covering the Azure fallback behavior

Testing

• uv run ruff check livekit-plugins/livekit-plugins-openai/livekit/plugins/openai/llm.py tests/test_config.py
• uv run pytest tests/test_config.py

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

nightcityblade seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.

[chenghao-mou]

We can simply use unknown as the last resort, as the model parameter is only used for metric/label.

[nightcityblade]

Addressed the review note in 4c0ede6. LLM.with_azure() now falls back to "unknown" when neither an explicit model nor an Azure deployment is provided, so we keep the metrics label neutral instead of defaulting to a specific model family. I also added a focused regression test for the no-model/no-deployment case.

Validation:

• uv run pytest tests/test_config.py
• uv run ruff check livekit-plugins/livekit-plugins-openai/livekit/plugins/openai/llm.py tests/test_config.py

[nightcityblade]

I pushed a follow-up commit to align with the review suggestion: Azure now keeps as the fallback model label unless the caller explicitly provides , and I updated the config expectation to match.\n\nI also rechecked the simple constructor path locally against the new behavior.

[nightcityblade]

I pushed a follow-up commit to align with the review suggestion: Azure now keeps "unknown" as the fallback model label unless the caller explicitly provides model, and I updated the config expectation to match.

I also rechecked the simple LLM.with_azure(...) constructor path locally against the new behavior.

[devin-ai-integration]

Devin Review found 1 new potential issue.

[devin-ai-integration]

🚩 Model value "unknown" is sent in the API request body to Azure

When model=None is passed to with_azure and no azure_deployment is provided either (as in test_llm_with_azure_uses_unknown_without_model_or_deployment), the string "unknown" will be sent as the model parameter in chat.completions.create() at livekit-agents/livekit/agents/inference/llm.py:351. With AsyncAzureOpenAI, if azure_deployment is None, the client may attempt to use the model value as the deployment name, which would cause a runtime 404 from the Azure API. This is arguably fine — the user failed to provide any routing info — but it's a change from the old default of "gpt-4o" which at least could match a real deployment name. The comment at llm.py:241-242 says the model is only used as a metric label, which is true when azure_deployment is set but not when it's absent.

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[Monkiia]

Good catch. The intent of the change was to avoid reporting a misleading default metric label when Azure callers provide a deployment but omit model, not to make the no-deployment path any more valid than it already is.

If maintainers want, I can follow up by narrowing the fallback so we only use "unknown" when azure_deployment is set, and otherwise keep the previous behavior for the invalid no-deployment case. That would preserve the neutral label improvement without changing the request-routing fallback.