The book has two parts. The first is a course in linear algebra, geometry, matrix decompositions, vector calculus, probability and optimisation. The second builds four familiar methods out of those pieces alone, to show that the same machinery sits under all of them.
What I took from it
The most valuable thing for my work is also the simplest: understanding what measuring similarity means. When a system says two documents are similar, it is computing an angle between two arrows in a space of many dimensions. Knowing that changes what I promise: the system finds what resembles in how words occur, not what is correct and not what is relevant to the question. It is a distinction a buyer cannot make alone and one worth telling them before signing.
The second is the idea that reducing dimensions is projecting: keeping the directions along which the data varies and discarding the rest. It reads directly onto indicator work, because most dashboards have far fewer real dimensions than they display.
The third is having lost my fear of the notation. After these pages a technical paper stops being a wall, and that alone is worth it in a market full of claims nobody verifies.
What I left out
A good deal, and I would rather be precise about it: I worked through the first part and the dimensionality chapter; the rest of the second part I went over without doing the exercises, which is another way of saying I did not learn it. It is a demanding book, it is available at no cost, and the distance between reading it and being able to use it is longer than the preface suggests.