I’m building a WhatsApp bot using LangGraph (with tools) that handles medical appointment scheduling. It works, but I keep running into an annoying behaviour time to time:
Human: hello do I have any appointment?
AI: Let me check if you have any future appointments. One moment, please!
The model keeps adding filler like “Let me check / One moment / Checking now…”
LLM_MODEL=gpt-4o-mini-2024-07-18
LLM_TEMPERATURE=0.2
I use Chat Completions API (not Assistant API).
I explicitly instruct it to execute tools silently and return only the final answer, but it keeps ignoring that.
Anyone found a reliable prompt pattern or other technique to stop these tool-narration phrases?