I built an AI assistant with OpenAI’s Assistants API and connected it to my Telegram and Instagram. I sell phone cases, and my clients message me through these platforms — my chatbot handles their questions.
Here’s the situation:
-
Client A started chatting with my assistant today. Everything worked fine — the assistant gave correct answers based on my knowledge base.
-
Tonight, I updated my knowledge base with new information about my products.
-
The next day, Client B came in and chatted. The assistant used the new knowledge base correctly.
-
But when Client A returned, the assistant gave wrong answers that seemed to come from the old knowledge base.
It looked like hallucination, but it was really the model pulling from older context. Basically, the assistant remembered or reused old info from before I updated the knowledge base. That’s why returning clients sometimes see outdated answers, while new clients only see the updated ones.