fix: strip think tags from content when reasoning=False

[yashranaway]

Description: Fix ChatOllama reasoning parameter not working as expected when set to False. When users explicitly disable reasoning by setting reasoning=False, thinking content (<think>...</think> tags) was still appearing in the response body. This fix adds a _strip_think_tags function that removes think tags from content when reasoning=False, ensuring clean responses without any reasoning content.

Issue: Fixes #33041

Dependencies: None - this is a pure bug fix that doesn't require any new dependencies.

Additional Details:

• Root Cause: The ChatOllama implementation correctly passed False to the Ollama API as the think parameter, but did not strip the <think> tags from the response content. The code only extracted reasoning content to additional_kwargs when reasoning=True, but had no logic to strip think tags when reasoning=False.

• Solution:

  • Added _strip_think_tags() function that removes <think>...</think> blocks using regex
  • Updated both sync (_iterate_over_stream) and async (_aiterate_over_stream) streaming methods to strip think tags when reasoning=False
  • Added proper whitespace cleanup to handle multiline content properly

• Behavior After Fix:

  • reasoning=True: Reasoning content extracted to additional_kwargs["reasoning_content"], main content clean
  • reasoning=False: Think tags stripped from main content, no reasoning content anywhere
  • reasoning=None: Default behavior (think tags remain in main content)

• Testing:

  • Created comprehensive unit tests for the _strip_think_tags function
  • Verified all existing unit tests still pass (30 passed, 2 skipped)
  • Comprehensive integration testing with mocked Ollama responses confirms the fix works correctly
  • Function correctly handles various edge cases including multiple think tags, empty content, etc.
  • Both invoke() and stream() methods properly strip think tags when reasoning=False
  • Backward compatibility verified - reasoning=True continues to work as expected

• Files Modified: libs/partners/ollama/langchain_ollama/chat_models.py (32 lines added)

This fix ensures that when users explicitly set reasoning=False, they get clean responses without any thinking content, which was the expected behavior that was missing.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Preview	Comments	Updated (UTC)
langchain	Ignored	Preview		Oct 2, 2025 2:24am

[codspeed-hq]

CodSpeed Instrumentation Performance Report

Merging #33042 will not alter performance

_{Comparing yashranaway:fix-chatollama-reasoning-false-bug (c83586c) with master (6f2d16e)}

Summary

✅ 1 untouched
⏩ 20 skipped¹

Footnotes

1. 20 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. ↩

[yashranaway]

hey @mdrxy, is there anything i need to update?

[yashranaway]

yup also #33116

[yashranaway]

👀

[mdrxy]

See comment on issue

[yashranaway]

@mdrxy done, sorry for that

[mdrxy]

#33041 (comment)