Training assistant with json and pdf data


I am trying to create a QnA bot with json and pdf data. I have uploaded all the knowledge after collating files within 20 number limit.

But the performance of Assistant is very less in comparison to same data and prompt used in GPT builder.

Is there any way to improve Assistant performance and make it similar to GPT builder ?

Also what can be best prompting technique to make the bot answer from the relevant document considering multiple documents have similar info ?