The JavaScript-based indexer lunr.js comes with is painfully slow and is unfit for large datasets. This is a native reimplementation that is mostly compatible.
It will use every .txt file in the working directory. It will not traverse subdirectories. It will write the results into index_v1.1.json file.
I highly recommend using a ramdisk as a working directory. Use the awesome ImDisk software if you are on Windows, or /dev/shm if you are on Linux.
The index produced does not use the save/load functionality integrated into lunr.js -- you don't need to modify lunr.js in any way but you still need quite a bit of custom external code: custom init, custom stemmer and more...