AzureChatOpenAI: deployment name missing from callback/tracing run data (ls_model_name defaults to gpt-3.5-turbo)

[fpfp100]

Description

When using AzureChatOpenAI from @langchain/openai, the Azure deployment name is not exposed in the LangChain callback/run data. run.extra.metadata.ls_model_name and run.extra.invocation_params.model both show "gpt-3.5-turbo" (the hardcoded default) instead of the actual deployment name.

This breaks all downstream observability/tracing tools (OpenTelemetry instrumentors, LangSmith, etc.) that rely on run metadata to identify the model.

Reproduction

import { AzureChatOpenAI } from "@langchain/openai";

const model = new AzureChatOpenAI({
  azureOpenAIApiDeploymentName: "gpt-4o-mini",
  azureOpenAIApiInstanceName: "my-instance",
  azureOpenAIApiKey: "...",
  azureOpenAIApiVersion: "2024-08-01-preview",
});

In a LangChain BaseTracer callback, run.extra for an LLM run shows:

run.extra.metadata.ls_model_name → "gpt-3.5-turbo"  (wrong)
run.extra.invocation_params.model → "gpt-3.5-turbo"  (wrong)
run.extra.invocation_params.azureOpenAIApiDeploymentName → undefined

The actual deployment name "gpt-4o-mini" is stored on this.azureOpenAIApiDeploymentName but never flows into the callback run data.

Root Cause

BaseChatOpenAI sets model = "gpt-3.5-turbo" as the default (line 35 of base.ts). When using AzureChatOpenAI with only azureOpenAIApiDeploymentName (and no explicit model), this.model stays as the default.

• getLsParams() sets ls_model_name: this.model → "gpt-3.5-turbo"
• invocationParams() returns { model: this.model, ... } → "gpt-3.5-turbo"
• Neither method is overridden by AzureChatOpenAI

Python parity

The Python AzureChatOpenAI already fixes this in two ways:

1. _identifying_params includes azure_deployment: self.deployment_name
2. _get_ls_params() falls back to self.deployment_name when self.model_name is None

# Python langchain_openai/chat_models/azure.py
@property
def _identifying_params(self):
    return {"azure_deployment": self.deployment_name, **super()._identifying_params}

See also: langchain-ai/langchain#24838 (Python equivalent)

Suggested Fix

Override getLsParams() in AzureChatOpenAI to fall back to the deployment name:

getLsParams(options) {
  const params = super.getLsParams(options);
  if (this.azureOpenAIApiDeploymentName && 
      (!params.ls_model_name || params.ls_model_name === "gpt-3.5-turbo")) {
    params.ls_model_name = this.azureOpenAIApiDeploymentName;
  }
  return params;
}

And/or override invocationParams() to include the deployment name:

invocationParams(options) {
  return {
    ...super.invocationParams(options),
    azureOpenAIApiDeploymentName: this.azureOpenAIApiDeploymentName,
  };
}

Impact

All callback-based instrumentation that reads model identity from LangChain run data is affected, including:

• OpenTelemetry distros (microsoft/opentelemetry-distro-javascript)
• LangSmith tracing
• Any custom BaseTracer implementations

Versions

• @langchain/openai: ^1.4.4
• langchain: ^1.2.31
• Node.js: 24.x

[obielin]

Hi! I'd like to pick this up — opening a PR shortly.

Plan, for visibility:

• Override getLsParams() in AzureChatOpenAI so ls_model_name falls back to
  azureOpenAIApiDeploymentName when no explicit model was passed (the
  BaseChatOpenAI default of "gpt-3.5-turbo" is what's leaking through today).
  An explicit model will still win.
• Override invocationParams() to also include azureOpenAIApiDeploymentName
  so callback/run metadata carries the deployment name regardless of whether
  model was set.
• Mirrors the Python parity fix in Error in trace, Trace for AzureChatOpenAI with gpt-4-turbo-2024-04-09 is not correct  langchain#24838.

Will add unit tests in
libs/providers/langchain-openai/src/azure/chat_models/tests/index.test.ts
covering the fallback, preservation of an explicit model, and the
deployment name showing up on invocationParams().

Could a maintainer assign this issue to me? Thanks!