Call ID:019c8b45-f9c4-7aa5-a9b3-05b3c2ca60a5 and another Call id:019c8b70-96a7-7551-93a8-42b763dbf416

Hey, After investigating there are the following problem we see: 1- ROOT CAUSE: Langfuse injects full KB context on EVERY turn Fix: Langfuse should only log/observe, it must NEVER inject context mid-call. Move all Langfuse writes to async post-call hooks. The LLM context should be managed solely by Vapi, not by a logging layer. 2- Anthropic 429 Rate Limit (4M tokens/min hit) Fix: Fix the Langfuse injection issue (Issue #1) this will immediately reduce token usage by ~80%. Also implement token-efficient context trimming: only send the last N turns + system prompt, not the entire history every time. 3- Google Gemini post-call analysis timed out (5s) Fix: Increase the post-call analysis timeout beyond 5s (10–15s is reasonable). Use a lighter model for structured data extraction