By using a vector database, we can bypass the finite context limitation of current LLMs and give the model only the context relevant to the task in hand.
I remember seeing that the model can get confused if you give it too much context.
especially if your asking it to do multiple things at one time.