Disclaimer: Experimental PR review

Greptile Summary

This PR modifies the usage parsing logic in the LangChain callback handler (CallbackHandler.py) to be more restrictive when filtering usage model values. The change on line 968 updates the filtering condition from not isinstance(v, str) to isinstance(v, int), meaning only integer values are now kept in the parsed usage model.

Based on the PR title referencing "qwen", this change appears to address an issue with Qwen models that return non-numeric string values in their usage metadata. The previous logic excluded only strings but allowed other types (including floats) to pass through. The new logic takes a more restrictive approach by only allowing integers.

This change integrates with Langfuse's usage tracking system, which uses Pydantic models like Usage and OpenAiUsage to structure token and cost data. These models typically accept optional integer fields for counts like input, output, total, and optional float fields for costs like input_cost, output_cost, total_cost. The filtering happens before this data reaches the Pydantic models.

Confidence score: 2/5

• This PR introduces potential data loss by filtering out legitimate float values that could represent costs or fractional usage metrics
• Score lowered due to overly restrictive filtering logic that may break existing functionality for models that return float-based usage data
• Pay close attention to langfuse/langchain/CallbackHandler.py line 968 and consider testing with various model providers that return float usage values