fix(bundles): expose model name on the Azure OpenAI chat component - #14574

Summary

Fixes #9855.

AzureChatOpenAIComponent (the "Azure OpenAI" model component) only exposed Deployment Name, with no way to tell langchain_openai.AzureChatOpenAI which underlying OpenAI model is behind that deployment. As reported in #9855, some Azure OpenAI configurations need the model name to resolve correctly, and it's also required for:

• Accurate token/cost tracing (AzureChatOpenAI._get_ls_params / _create_chat_result fall back to the deployment name when model_name is None, which usually doesn't match a real model id).
• langchain_openai's model-name-gated feature checks (e.g. the parallel_tool_calls support check keys off model_name == "gpt-4o", and reasoning-specific params for o-series/gpt-5 models).

Note: for plain Chat Completions routing, azure_deployment alone is already enough to reach the right deployment (the endpoint's base_url is built from it at client construction), so the change here doesn't affect routing for the common case — it closes the actual gap in the issue (no way to declare the model at all) and the secondary correctness issues above.

Changes

• Added an optional, advanced model input to AzureChatOpenAIComponent, matching the field name already used by the sibling AzureOpenAIEmbeddingsComponent in the same file.
• Passed it through as AzureChatOpenAI(model=self.model or None, ...).
• When left blank, model=None is sent — identical to current behavior, so existing saved flows are unaffected.

Kept deliberately minimal: no default value is forced on the new field (unlike the embeddings sibling / OpenAI component), because a forced default could silently mismatch a user's actual deployed model for existing flows that don't set this field.

One naming detail worth calling out for review: I initially named the field model_name (matching OpenAIModelComponent), but that collides with lfx/base/models/model_input_constants.py's shared "unified model providers" registry, which special-cases any input literally named model_name and expects a DropdownInput with .combobox support. Since AzureChatOpenAIComponent is already registered there (currently is_active: False), that name broke import of model_input_constants.py itself (and everything that imports it — altk, cuga, the agent component, etc.) with a pydantic ValidationError. Renaming the field to model avoids that collision entirely and still matches the sibling embeddings component's convention.

Test plan

• Added src/backend/tests/unit/components/bundles/azure/test_azure_openai_component.py covering:
  • Input structure (the new field exists, is optional and advanced).
  • build_model() passes model= through to AzureChatOpenAI.
  • Omitting the field preserves prior behavior (model=None).
  • Exception handling is preserved.
  • The inherited test_latest_version smoke test.
• Ran the full suite to confirm no regressions:
  • uv run pytest src/backend/tests/unit/components/bundles/azure/ — 7 passed, 4 skipped.
  • uv run pytest src/backend/tests/unit/components/bundles/ src/backend/tests/unit/base/models/ src/backend/tests/unit/test_load_components.py — 355 passed, 94 skipped, 3 pre-existing failures in altk unrelated to this change (verified they fail identically on unmodified main, environment/API-key related).
  • uv run --package lfx pytest src/lfx/tests/unit/base/models/ — 329 passed (covers the shared unified-model-providers registry that the field-name collision above would have broken).
  • ruff check / ruff format --check on both changed files — clean.

Summary by CodeRabbit

• New Features

  • Added an advanced model configuration option for Azure OpenAI connections.
  • The selected model is now applied when creating the chat model.

• Bug Fixes

  • Improved handling of empty model settings and connection errors with clearer contextual messages.

• Tests

  • Added coverage for default setup, input definitions, model configuration, argument forwarding, and error handling.