From How To Geek:
Microsoft Recall is a new, special kind of AI-powered search [...] If you were writing a short story about a cat and forgot where you saved it, you should be able to ask “Where did I save my story about the cat” and it’ll be able to tell you. [...] Recall is one of several new AI features that are going to require a Neural Processing Unit (NPU), which is a special kind of processor that has been optimized for machine learning and artificial intelligence operations. [...] Plenty of current devices have NPUs integrated into their CPUs, but their performance generally falls short of the magic 40 trillion operations per second (TOPs) we’ve seen thrown around as a minimum requirement. [...] We don’t know just how many of these AI tasks can be offloaded to the GPU in the event that you don’t have a NPU that is up to the job.
typing "where did i save the cat story" into my £2000 PC and listening to the fans scream as GPU utilisation jumps to 100% for five minutes so it can say "The file you’re looking for, cat.docx, which contains the story you wrote about your cat, is located on your Desktop. To access it, navigate to your Desktop folder and you should find cat.docx there. However, as a large language model it's important to keep in mind that I can produce output that may be inaccurate, so please ensure to seek reputable sources for any information."
Or, well, maybe you could use a small neural network, a kind that would be capable of running on a CPU and associate keywords like "story" with file types.
And then just have a file crawler (or Word itself) index your file, make a tag tree with top-100 most popular searchable words (as in, don't count articles, adpositions, conjunctions, etc) and store them in a database somewhere. That is it.
It has been done before.
It has been done by Google.
It has been done by OSs.
You do not need LLMs for this.
Specifically, there's a type of AI model called an embedding model that maps arbitrary text onto a vector space and then you can use vector similarity to find the N nearest vectors and then retrieve the original context. It's actually an extremely effective and general solution to the problem and it's capable of making really uncanny/relevant associations.
I would actually go so far as to say that embedding models are the real breakthrough technology in the field of AI (even more so than completion/chat models).