fix(openai): Improve vLLM API compatibility and error messages

[KaparthyReddy]

Description

Fixes #32252 - Resolves the issue where LangChain-OpenAI throws TypeError: Received response with null value for choices when using vLLM's OpenAI-compatible API.

Problem

When using vLLM's OpenAI-compatible API through LangChain, the response parsing would fail with a null choices error, even though the choices field exists in the raw response. This was caused by how the OpenAI SDK's model_dump() method handles non-OpenAI API responses.

Solution

This PR improves response handling in the _create_chat_result method:

1. Robust response serialization: Added fallback logic that tries model_dump_json() if model_dump() fails
2. Better error messages: Enhanced the null choices error to include:
   • Available response keys
   • Response object type
   • Hints about API compatibility issues
3. Comprehensive tests: Added tests for:
   • vLLM-style responses with valid choices
   • Improved error message content

Changes

• Modified langchain_openai/chat_models/base.py:
  • Added try-except block with fallback serialization (lines 1271-1287)
  • Improved error message with debugging info (lines 1302-1309)
• Added tests in tests/unit_tests/chat_models/test_base.py:
  • test_vllm_response_with_valid_choices() - Verifies vLLM responses work
  • test_improved_null_choices_error_message() - Validates error message improvements

Testing

All existing tests pass (279 passed, 4 xfailed, 1 xpassed).

New tests verify:

• vLLM-style responses are handled correctly
• Error messages provide useful debugging information

Usage Example

After this fix, users can successfully use vLLM endpoints:

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    api_key="your-runpod-key",
    base_url="https://api.runpod.ai/v2/<endpoint-id>/openai/v1",
    model="your-model"
)

# This will now work without null choices error
response = llm.invoke("Hello!")

Resolves #32252

[codspeed-hq]

CodSpeed Performance Report

Merging #33662 will not alter performance

_{Comparing KaparthyReddy:fix/vllm-null-choices-error (663d485) with master (f174295)}

Summary

✅ 6 untouched
⏩ 28 skipped¹

Footnotes

1. 28 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩