Langchain model token counting bug fix

[adriantqw]

For models with reasoning tokens, the _parse_usage function in the Langchain CallbackHandler does not successfully parse token usage metadata to PostHog.

[greptile-apps]

PR Summary

Fixed token counting for language models with reasoning tokens by enhancing the _parse_usage function in posthog/ai/langchain/callbacks.py to properly handle usage metadata parsing.

• Added fallback paths to check usage information in both generation_info and message properties of model responses
• Implemented default empty dictionary fallback if neither usage source exists
• Critical fix ensures accurate token counting metrics in PostHog analytics, particularly for models using reasoning tokens

_{1 file reviewed, no comments
Edit PR Review Bot Settings | Greptile}