We are facing inconsistent behavior with our VAPI-based AI assistant, currently using the GPT-4.1 model. The assistant performs well in some cases but often misinterprets clear user input or fails to handle context accurately in similar conversations. Issue Details: • Call ID: 43e534c1-9001-4f8a-88b0-a434c0c3534b • Timestamp of Issue Start: Around 00:19 • Expected Behavior: • When the user says: “My last name is Tejani”, the assistant should correctly capture and confirm “Tejani” as the last name. • When the user asks: “Please check if I have any appointment today”, the assistant should interpret “today” as the current date and respond accordingly. • Valid mobile numbers should be accepted without false validation errors. Actual Behavior: • The assistant misinterprets “Tejani” as “Tejantu” inconsistently across different calls. • For date-related questions, it sometimes interprets “today” as a past date. • Mobile number validation sometimes fails even when the format and data are correct. These issues are impacting the user experience and causing concern with our client as we approach production release. We would greatly appreciate any insight or guidance on how to resolve these inconsistencies or improve prompt engineering/context handling for more reliable performance.

Improve your prompts by adding explicit instructions and clear context anchors, adjust model parameters like temperature for consistency, use Vapi’s call logs to identify and troubleshoot errors, and consider upgrading to the GPT-4o model for more reliable performance.