There are two limitations for your idea.

The first is computing power. Language transformers (like GPT-3 and 4) require a beefy hardware to run. As an example, to run one of the open source transformers (GPT-NeoX) locally, you need at least two high end graphics cards. We are still a long way from running this on a phone - but we can run the language transformer on a server and have your phone talk to the server.

The second limitation is result accuracy. As someone else noted, language transformers occasionally hallucinate wrong information. Think about it this way: What a language transformer does is basically that it strings together words in the most plausible way it can. It doesn't understand the words. For simple problems, the most plausible sounding solution just also happens to be factually correct. For more complex problems, it will hallucinate plausible sounding, but factually wrong information.

There are two ways to tackle this problem.

The first is the one you've mentioned, fine-tuning the language model on established data like textbooks. This will reduce hallucinations, but likely not by a great amount.

The second solution is akin to what Bing currently does: Combine the language transformer with regular full-text search. You ask your phone "What's the max fill on a cable tray type XY?", this gets run through the language transformer to extract key search terms from your question, in this case "max fill" and "cable tray XY". A regular text search for these key words is performed and the system finds a couple possibly relevant hits: Page 12 paragraph 10, page 33 paragraph 2 and so on. These relevant paragraphs are then fed into the language transformer together with your original query: "Hey language transformer, asnwer the question Whats the max fill on cable tray XY using information from these paragraphs: ..."

The language transformer then summarizes the paragraphs in natural language according to your query, spits out a result and your phone tells you "The max fill is 42 cables. Source: Textbook Name Page 12."

So yes, this is very possible and personally I expect to see this popping up a lot in the future.

You can do that very thing already with Bing, it will search the web and give you a source for its result.