This is great!

Some far-reaching implementation changes, I'll set some time aside to read over them later this week.

Cheers.

I'm not 100% sure I chose the correct approach, and I think there may be a simpler approach possible. But I didn't fully understand the codebase when I started, so this is what ended up happening. One thing that confused me was the use of the atom variable name to mean both a SearchAtom and a single word, so I changed instances of the latter to "token".

Hi there.

"This week" has turned into "This year".

The "do not match quoted phrases across field boundaries" commit is great, I have some tests for this already, hopefully this will satisfy those.

I had a play with implementing stemmer support today, first on my own, then I read through your implementation. While I can see the usefuleness of it, it does add a lot of complexity at all levels of the library. It also has quite an effect on index sizes. I'll probably leave this for just now, though I am tempted to come back to it in the future.