This reminded me of another, related task that comes up a lot in machine learning and IR: searching by vectors (rather than keywords). Aka k-NN (k nearest neighbors) search.

Just as with fulltext search, there's no shortage of clever approximate algorithms for kNN vector search (with the additional caveat of the "curse of dimensionality").

And just as here, without careful tuning, a well-optimized brute force search blows most "clever", widely popular kNN implementations out of the water on large datasets [1]. There's certainly something to be said for simplicity and good "systems engineering".

[1] http://rare-technologies.com/performance-shootout-of-nearest...