How to restrict OpenAI Assistant(GPT) dependency on internet for response?
I am using gpt-3.5 turbo. Uploaded documents to storage. Getting response from Internet instead uploaded content, when prompts are executed.
Welcome to the Dev community!
The model won’t be accessing the internet since this currently isn’t a feature in the API, more likely it’s simply not retrieving the information in the documents and relying on its training instead.
Perhaps you could try referencing specific documents for it to use in the system prompt? Either way this is either a prompting issue or your docs aren’t well suited to RAG.
1 Like
Thanks for quick response.
I have the grounded data as 3 text files.
I expect GPT model to look in to the grounded data only. Is there an attribute in request that explicitly mentions GPT model to look at grounded data only.
(In Azure OpenAI, there is an attribute “Scope:True”, that ensures this)
I highly recommend reading through this page in the documentation if you haven’t already, it might answer some of your questions in regards to enforcing the tool use.
My preferred solution if its only 3 documents is to explicitly mention the document names in the system prompt, play around with this enough and you can usually get it working. You might want to try with gpt-4o-mini too as it’s cheaper than even 3.5.