In google search you can use symbols and words to make the results more precise, tokens like: after:, before:, site:, imagesize:, filetype: and many more.
GPT-style models with a well crafted prompt can be used to have a compiler of sorts for the natural language query, which will be turned into a google-specific query that is more precise and close to what the user is most likely looking for.
And if we tell the “compiler” things like: the user current date and timezone, geographic region—and other info the user might care about, it can take these into account in the final compiled query.
For example, searching on google for: pie recipes without milk. from reddit in the last 42 hours returns:
The same query ran with our demo script gooogle—like google but with one more humpf–outputs the following compiled query: pie recipes -milk site:reddit.com after:2022-10-07 before:2022-10-09 and searching this query on google returns:
As we can see, we got better results with the same search query for what we were actually looking for–posts no older than 42 hours, from reddit, without milk in this case. With more fine-tuning there’s a lot more that can be accomplished without having to know and remember a myriad of magic keywords.
A prototype can be found here.