why

LiteLLM's synchronous completion() method was blocking the event loop in async handlers, preventing concurrent execution of multiple LLM calls. This caused performance degradation when multiple operations needed to run in parallel.

what changed

• Converted LLMClient.create_response() from sync to async method using litellm.acompletion()
• Updated inference.observe() and inference.extract() functions to be async
• Modified all handlers (ObserveHandler, ExtractHandler) to await async inference calls
• Updated mock LLM client's create_response() method to be async for test compatibility

test plan

• run CI tests and inspect that things are working as expected