User input
So that’s where the user inputs. It’s where you input your query because you’re trying to find the best pair of running shoes for you personally, not just what ChatGPT’s browser thinks is the best pair of running shoes.
Context understanding
So then what happens is there’s a context understanding that goes on. It analyzes what your current browsing session is, and that could be multiple tabs.
Now just so you know, the model can see what you’re browsing only with your permission. So you can switch this on or off in ChatGPT’s Atlas at the moment.
Information retrieval
Then what happens is there’s an information retrieval. If ChatGPT feels like it already has enough information in its local context, it can use that to answer your query directly. But more often than not, what it might do is use its retrieval APIs in order to search for newer, up-to-date information to give you a more complete and better answer, specifically to what you’ve asked.
Reasoning and answer generation
Then it moves into reasoning and answer generation.
The LLM fuses multiple things — the page context of what you’ve been looking at on the browser tabs in ChatGPT Atlas, the user intent, what it thinks you’re actually asking for to try and give you the answer that is most relevant to you, the retrieval of the supplied sources if it had to use the API to search live on the internet for answers, and relevant memory.
So it can look at your past chats and conversations that you’ve had in ChatGPT and decide if they’re relevant to also add into its answer.
