fix(langchain): resolve request model from ls_model_name metadata for LangGraph tool spans

[leorivastech]

When you run a LangChain chat model through LangGraph with a provider that
doesn't surface the model in its serialized kwargs — for example
init_chat_model(..., model_provider="bedrock_converse") bound with tools — the
resulting spans come out with gen_ai.request.model and gen_ai.response.model
set to unknown. The span dump in #3098 shows it nicely: everything else is
there (prompts, tool calls), but the model name is gone.

I traced it down to set_request_params. It looks for the model in two spots:
the callback's invocation_params, and serialized["kwargs"] (via
set_chat_request). For something like ChatOpenAI the name lives in
serialized["kwargs"]["model_name"], so it resolves fine and nobody notices.
But for LangGraph tool-orchestrated models neither spot has it, so we just give
up and write "unknown".

The thing is, LangChain does hand us the model name — it's in the callback
metadata as ls_model_name, which _get_ls_params fills in automatically for
every chat model. We were already receiving that metadata in
on_chat_model_start and passing it to _create_llm_span, just never down to
set_request_params.

So the fix is small: thread the metadata through and, when the usual lookups come
up empty, fall back to ls_model_name before defaulting to unknown. It's the
same key we already trust on the response side (_extract_model_name_from_association_metadata,
from #3237), so this just brings the request side in line. When the model is
resolvable the normal way, nothing changes — it's purely a fallback.

For testing I added tests/test_request_model_extraction.py. It drives the
callback handler directly against an in-memory exporter, so it needs no cassettes
and no API keys: one test reproduces the LangGraph/bedrock_converse case (it fails
on main with assert 'unknown' == 'us.anthropic...' and passes here), and
another makes sure an explicit model in serialized["kwargs"] still takes
precedence over the metadata fallback. The full langchain suite stays green (151
passed) and ruff is clean.

Fixes #3098

---

• I have added tests that cover my changes.
• If adding a new instrumentation or changing an existing one, I've added screenshots from some observability platform showing the change. — this is an attribute-mapping fix, not a UI/instrumentation addition, so I went with a span-asserting test instead of screenshots.
• PR name follows conventional commits format.
• (If applicable) I have updated the documentation accordingly.

Summary by CodeRabbit

• Improvements

  • Enhanced model identification in OpenTelemetry traces for tool-orchestrated LangChain chat models by supporting metadata-based model attribution when model information is not available through default resolution paths.

• Tests

  • Added test coverage for model extraction logic in tracing scenarios.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: 339cc9d9-9810-4570-8924-89b7d6dec8b0

📥 Commits

Reviewing files that changed from the base of the PR and between fc33f1c and b7d941d.

📒 Files selected for processing (3)

• packages/opentelemetry-instrumentation-langchain/opentelemetry/instrumentation/langchain/callback_handler.py
• packages/opentelemetry-instrumentation-langchain/opentelemetry/instrumentation/langchain/span_utils.py
• packages/opentelemetry-instrumentation-langchain/tests/test_request_model_extraction.py

---

📝 Walkthrough

Walkthrough

Adds an optional metadata parameter to set_request_params, set_llm_request, and set_chat_request in span_utils.py. When model derivation from kwargs/invocation_params fails, the code falls back to metadata["ls_model_name"]. The callback_handler.py threads the existing metadata argument through both on_chat_model_start and on_llm_start. Two unit tests cover the fallback and serialized-payload precedence behaviors.

Changes

Metadata-based model name fallback

Layer / File(s)	Summary
Model name fallback in span_utils packages/opentelemetry-instrumentation-langchain/opentelemetry/instrumentation/langchain/span_utils.py	set_request_params gains an optional metadata parameter and reads ls_model_name from it when kwargs/invocation_params yield no model name; set_llm_request and set_chat_request expose the same optional metadata parameter and forward it.
metadata threading in callback_handler packages/opentelemetry-instrumentation-langchain/opentelemetry/instrumentation/langchain/callback_handler.py	on_chat_model_start and on_llm_start now pass the existing metadata argument to set_request_params, set_chat_request, and set_llm_request.
Regression and precedence tests packages/opentelemetry-instrumentation-langchain/tests/test_request_model_extraction.py	New test file drives on_chat_model_start + on_llm_end against an InMemorySpanExporter-backed handler and asserts that gen_ai.request.model resolves from metadata["ls_model_name"] when invocation_params are empty, and that serialized["kwargs"]["model_name"] takes precedence when present.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

Poem

🐇 A model named "unknown" hid in the trace,
No kwargs or params to show its face.
But metadata whispered, "ls_model_name here!"
The rabbit patched it, the span became clear.
Now LangGraph spans bloom with names bright and true —
No more mystery models, hurrah and woo-hoo! 🎉

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 61.54% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The PR title clearly and specifically describes the main change: resolving the request model from ls_model_name metadata for LangGraph tool spans, which directly addresses the core issue.
Linked Issues check	✅ Passed	The PR fully addresses issue #3098 by implementing a metadata fallback mechanism to populate gen_ai.request.model and gen_ai.response.model with the actual model name for LangGraph-orchestrated chat models instead of defaulting to unknown.
Out of Scope Changes check	✅ Passed	All changes are directly scoped to resolving the reported issue: threading metadata through callback handler and span utilities, adding model resolution fallback logic, and including comprehensive test cases with no unrelated modifications.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}