feat(langchain): migrate ChatYandexGPT to Chat API

[temrjan]

Summary

Migrates the LangChain integration from the deprecated GRPC Foundation Models API to the modern Chat API (/_chat/completions/), enabling tool call support in ChatYandexGPT.

• New ChatYandexGPT in _chat/completions/langchain.py using HTTP REST
• Parse tool_calls from Chat API response into AIMessage.tool_calls
• Support streaming with tool_call_chunks
• Add .langchain() method to BaseChatModel
• Keep langchain as optional dependency
• Old _models/completions/langchain.py untouched

Changes

File	Change
_chat/completions/langchain.py	New — ChatYandexGPT adapter for Chat API
_chat/completions/model.py	+11 lines — .langchain() method
tests/langchain_/test_chat_api.py	New — 7 tests
examples/langchain/chat/chat_api_invoke.py	New — usage example

Design decisions

• _transform_messages() converts LangChain messages → Chat API dicts, then existing messages_to_json() handles serialization
• _parse_tool_calls() filters out malformed tool calls (tc.function is None)
• Streaming uses tool_call_chunks with args as JSON string (LangChain convention)
• Sync methods delegate to async via run_sync_impl (same pattern as legacy adapter)
• Multimodal content supported via list-type HumanMessage.content

Test plan

• test_ainvoke / test_invoke — basic completion
• test_astream / test_stream — streaming
• test_usage_metadata — token usage stats
• test_response_metadata — finish_reason, model

Next

This PR is the foundation for a follow-up PR adding bind_tools() support (LangChain tools → Yandex FunctionTool conversion + full round-trip).

Refs #142

🤖 Generated with Claude Code

[vhaldemar]

I mean, it is not going to work, because we have no network access at tests.
We usually using pytest-recording for tests like this, recording cassetetes with network responses/requests

[vhaldemar]

I remembered one more thing about our langchain support: we are restricting langchain-core dependency version, because new lanchain have new pydantic dependency, which is breaking backward compatibility.

Probably it have to be fixed if we want to bring this integration back to living world :D

But anyway, this is not about this PR.

[vhaldemar]

I'm kinda feel that claude didn't understand some contexts here.

I mean, HTTP /_chat/completions/ is means nothing for ordinary user who will read this doc

[vhaldemar]

This is unnecessary info for example :)

[temrjan]

Thanks for the review! Pushed fixes:

Tests — switched to pytest-recording (VCR cassettes), matching the test_completions.py pattern:

• pytestmark = [pytest.mark.asyncio, pytest.mark.vcr, pytest.mark.require_env('langchain_core')]
• async_sdk only (no parametrization)
• 2 focused tests: test_ainvoke (with metadata assertions) + test_astream

Docstrings — removed implementation details (/_chat/completions/, GRPC mentions) from module, class, and example docstrings.

Cassettes — I don't have API access to record them. Could you help record the VCR cassettes, or point me to a test folder_id/API key I could use?

pytest tests/langchain_/test_chat_api.py --record-mode=once --env langchain_core -v

[vhaldemar]

Cassettes — I don't have API access to record them. Could you help record the VCR cassettes, or point me to a test folder_id/API key I could use?

I'm not sure, I will try to ask my collegues, if it is posibble to give you some kind of grant to use yandex cloud.

[vhaldemar]

Hey!

1. You should get an yandex cloud account with payment method setupped
2. Create a support ticket, something like "I'm asking for the grant for work with feat(langchain): migrate ChatYandexGPT to Chat API #208"
3. Post here an ID of support ticket

[temrjan]

Hey @vhaldemar! Thanks for the guidance.

Done — created a support ticket: KF218173

Let me know if anything else is needed from my side!

[vhaldemar]

Hey, collegues told me your ticket is done.

Also, please? make a rebase or merge to master to fix docs build issue.

[temrjan]

Hey @vhaldemar! Rebased on master, docs build should be fixed now. Cassettes are recorded with real API. Ready for another review when you have time!